---

Documentation and Governance

"If a dataset cannot explain where it came from, it cannot explain what it taught."

Overview

Documentation and governance convert a data pipeline into a reproducible, reviewable, and accountable artifact. In an LLM training run, data is not an inert pile of text; it is the empirical distribution that defines the examples, losses, risks, and capabilities the model will see.

This section is written as LaTeX Markdown. Inline mathematics uses $...$, and display equations use $$...$$. The goal is to connect data engineering decisions to mathematical objects such as records $r_i$, token sequences $x_{1:T}$, filters $f(x)$, hashes $h(x)$, mixture weights $\boldsymbol{\alpha}$, and empirical expectations.

The scope is deliberately narrow: this chapter owns the training-data pipeline. Tokenizer design, GPU training systems, benchmark methodology, alignment objectives, and production MLOps each have their own canonical chapters. Here we study the data objects that those later systems consume.

Prerequisites

• Contamination and Dedup Audits

Learning Objectives

After completing this section, you will be able to:

• Define data cards, provenance graphs, license fields, risk registers, and release checklists
• Write dataset documentation that explains intent, sources, processing, and limitations
• Represent lineage through source URIs, hashes, transforms, and version pins
• Design governance controls for access, PII review, takedown, and release approval
• Track dataset versions and diff reports
• Link trained models back to exact data manifests
• Explain why documentation is a user-facing product, not a README afterthought
• Prepare evidence needed for reproducibility and responsible release

---

---

1. Intuition

Intuition gives the conceptual and mathematical layer for documentation and governance. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

1.1 A dataset without documentation is not reproducible

A dataset without documentation is not reproducible is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection

$\mathcal{D} = \{r_i\}_{i=1}^n$

with record-level metadata and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For data card, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

1.2 Governance as risk control

Governance as risk control is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For provenance, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

1.3 Dataset users as stakeholders

Dataset users as stakeholders is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For lineage, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

1.4 Responsible release

Responsible release is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For license, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

1.5 Data cards and data statements

Data cards and data statements is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For risk register, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

2. Formal Definitions

Formal Definitions gives the conceptual and mathematical layer for documentation and governance. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

2.1 Dataset card

Dataset card is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record- level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For data card, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

2.2 Provenance graph

Provenance graph is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For provenance, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

2.3 License vector

License vector is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record- level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For lineage, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

2.4 Consent/permission field

Consent/permission field is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For license, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

2.5 Risk register

Risk register is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record- level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For risk register, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

3. Dataset Documentation

Dataset Documentation gives the conceptual and mathematical layer for documentation and governance. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

3.1 Intended use

Intended use is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record- level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For data card, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

3.2 Collection process

Collection process is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For provenance, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

3.3 Processing pipeline

Processing pipeline is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For lineage, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

3.4 Known limitations

Known limitations is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For license, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

3.5 Evaluation and audit results

Evaluation and audit results is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For risk register, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

4. Provenance and Lineage

Provenance and Lineage gives the conceptual and mathematical layer for documentation and governance. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

4.1 Source URI

Source URI is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For data card, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

4.2 Snapshot time

Snapshot time is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record- level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For provenance, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

4.3 Transform history

Transform history is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For lineage, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

4.4 Hash chain

Hash chain is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For license, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

4.5 Rebuild command

Rebuild command is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record- level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For risk register, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

5. Governance Controls

Governance Controls gives the conceptual and mathematical layer for documentation and governance. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

5.1 Access control

Access control is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record- level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For data card, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

5.2 License compatibility

License compatibility is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For provenance, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

5.3 PII review

PII review is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For lineage, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

5.4 Takedown/removal workflow

Takedown/removal workflow is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For license, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

5.5 Release approval checklist

Release approval checklist is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For risk register, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

6. Dataset Versioning

Dataset Versioning gives the conceptual and mathematical layer for documentation and governance. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

6.1 Semantic dataset versions

Semantic dataset versions is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For data card, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

6.2 Diff reports

Diff reports is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record- level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For provenance, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

6.3 Deprecation

Deprecation is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For lineage, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

6.4 Reproducible manifests

Reproducible manifests is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For license, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

6.5 Model-to-data linkage

Model-to-data linkage is part of the canonical scope of documentation and governance. We model the relevant object as a finite collection $\mathcal{D} = \{r_i\}_{i=1}^n$ with record-level metadata $m_i$ and text or token content $x_i$. The practical question is whether the transformation preserves the intended empirical distribution.

A useful local invariant is:

$$ \text{valid}(r_i, \mathcal{S}) = 1 \quad \Longrightarrow \quad r_i \text{ can be consumed by the next pipeline stage.} $$

For risk register, the invariant should be explicit enough that a checker can fail fast. If the invariant is only written in a notebook comment or an engineer's memory, it will not protect a long-running data build.

Examples: - A small local experiment can store this object in memory; a frontier-scale run must store it as sharded, versioned, validated records. - The mathematical object is simple, but the operational contract must survive restarts, parallel workers, schema changes, and audits. - The notebook for this section uses synthetic data so the same ideas can be executed without external files.

Non-examples: - A path on disk without a manifest is not a reproducible dataset. - A metric dashboard without record-level lineage is not a provenance system. - A filter threshold without an audit sample is not evidence of quality.

Implementation consequence: every transformation should report both a count and a rate. If $n_{\mathrm{in}}$ records enter the stage and $n_{\mathrm{out}}$ records leave, the acceptance rate is

$$ a = \frac{n_{\mathrm{out}}}{n_{\mathrm{in}}}. $$

A sudden change in $a$ is a data-drift signal even when the code still runs. This is why pipeline math is inseparable from logging, manifests, and audit slices.

For LLM work, the token-weighted view is often more important than the document-weighted view. A filter that removes 5 percent of documents may remove 30 percent of tokens if it targets long documents. The corresponding token acceptance rate is

$$ a_{\mathrm{tok}} = \frac{\sum_i f(r_i)\,T_i}{\sum_i T_i}, $$

where $T_i$ is the token count or a deterministic token-count estimate. The distinction matters for compute budgets, mixture proportions, and scaling-law interpretation.

7. Common Mistakes

#	Mistake	Why It Is Wrong	Fix
1	Trusting a file because it exists	A zero-byte or unparsable artifact can still pass a loose path check	Validate content and parseability
2	Counting documents but not tokens	Long documents dominate compute	Report both document and token rates
3	Changing schemas without versioning	Old and new records become indistinguishable	Pin schema versions in every record
4	Dropping metadata during transforms	Audits and removals become impossible	Preserve source and transform lineage
5	Using nondeterministic ordering	Rebuilds cannot be compared	Seed and record ordering rules
6	Ignoring failed records	Silent loss can bias the corpus	Quarantine and summarize failures
7	Treating filters as neutral	Filters encode preferences and tradeoffs	Ablate and audit every major filter
8	Mixing train and eval sources	Evaluation becomes contaminated	Run overlap audits before release
9	Optimizing one aggregate score	Small domains can regress	Track slice metrics
10	Skipping data cards	Users cannot judge intended use or risk	Publish structured documentation
11	Assuming licenses are uniform	Source terms can conflict	Track license at source and record level
12	Forgetting reproducible manifests	The same name can refer to different data	Use hashes and version pins

8. Exercises

1. (*) Build a synthetic data card example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
2. (*) Build a synthetic provenance example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
3. (*) Build a synthetic lineage example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
4. (**) Build a synthetic license example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
5. (**) Build a synthetic risk register example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
6. (**) Build a synthetic version example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
7. (**) Build a synthetic governance example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
8. (***) Build a synthetic data card example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
9. (***) Build a synthetic provenance example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
10. (***) Build a synthetic lineage example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.

9. Why This Matters for AI

Concept	AI impact
data card	Controls what examples, gradients, risks, or audits the model pipeline can represent
provenance	Controls what examples, gradients, risks, or audits the model pipeline can represent
lineage	Controls what examples, gradients, risks, or audits the model pipeline can represent
license	Controls what examples, gradients, risks, or audits the model pipeline can represent
risk register	Controls what examples, gradients, risks, or audits the model pipeline can represent
version	Controls what examples, gradients, risks, or audits the model pipeline can represent
governance	Controls what examples, gradients, risks, or audits the model pipeline can represent

Data pipeline quality is model quality in delayed form. The model eventually converts these records into gradients; any unresolved ambiguity becomes either wasted compute, misleading evaluation, memorization risk, or irreproducible science.

10. Conceptual Bridge

This section connects the previous and next pieces of the curriculum as follows:

raw sources -> records -> validation -> assembly -> audits -> documentation -> mixture

The next section is Data Mixture Optimization. It uses the contracts established here and moves one step further through the LLM data pipeline.

References

• DataComp-LM
• FineWeb
• Dolma
• WIMBD
• Deduplicating Training Data Makes Language Models Better
• DoReMi
• Data Cards
• A Pretrainer's Guide to Training Data
• Hugging Face Datasets loading docs

$m_i$