Lesson overview | Previous part | Lesson overview

Data Format Standards: Part 4: Storage Formats to References

4. Storage Formats

Storage Formats gives the conceptual and mathematical layer for data format standards. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

4.1 JSONL

JSONL is part of the canonical scope of data format standards. 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 record, 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 Parquet and Arrow

Parquet and Arrow is part of the canonical scope of data format standards. 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 schema, 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 Tokenized binary arrays

Tokenized binary arrays is part of the canonical scope of data format standards. 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 JSONL, 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 Sharded compressed files

Sharded compressed files is part of the canonical scope of data format standards. 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 metadata, 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 Manifest files and checksums

Manifest files and checksums is part of the canonical scope of data format standards. 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. Validation Rules

Validation Rules gives the conceptual and mathematical layer for data format standards. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

5.1 Required keys

Required keys is part of the canonical scope of data format standards. 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 record, 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 Type validation

Type validation is part of the canonical scope of data format standards. 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 schema, 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 Unicode and whitespace normalization

Unicode and whitespace normalization is part of the canonical scope of data format standards. 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 JSONL, 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 Text-length constraints

Text-length constraints is part of the canonical scope of data format standards. 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 metadata, 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 Deterministic IDs and hashes

Deterministic IDs and hashes is part of the canonical scope of data format standards. 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. Applications

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

6.1 Pretraining corpora

Pretraining corpora is part of the canonical scope of data format standards. 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 record, 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 Continual pretraining

Continual pretraining is part of the canonical scope of data format standards. 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 schema, 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 SFT

SFT is part of the canonical scope of data format standards. 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 JSONL, 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 Preference data

Preference data is part of the canonical scope of data format standards. 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 metadata, 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 Data release packages

Data release packages is part of the canonical scope of data format standards. 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.

7. Common Mistakes

8. Exercises

1. (*) Build a synthetic record example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
2. (*) Build a synthetic schema example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
3. (*) Build a synthetic JSONL example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
4. (**) Build a synthetic metadata example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
5. (**) Build a synthetic provenance example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
6. (**) Build a synthetic token sequence example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
7. (**) Build a synthetic manifest example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
8. (***) Build a synthetic record example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
9. (***) Build a synthetic schema example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
10. (***) Build a synthetic JSONL example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.

9. Why This Matters for AI

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 JSONL Generation. 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