Lesson overview | Previous part | Lesson overview

Full Dataset Assembly: Part 4: Assembly Algorithms to References

4. Assembly Algorithms

Assembly Algorithms gives the conceptual and mathematical layer for full dataset assembly. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

4.1 Concatenation

Concatenation is part of the canonical scope of full dataset assembly. 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 source set, 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 Weighted sampling

Weighted sampling is part of the canonical scope of full dataset assembly. 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 mixture weight, 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 Stratified sampling

Stratified sampling is part of the canonical scope of full dataset assembly. 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 token budget, 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 Train/validation/test split

Train/validation/test split is part of the canonical scope of full dataset assembly. 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 manifest, 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 Deterministic shuffling

Deterministic shuffling is part of the canonical scope of full dataset assembly. 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 shard, 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. Tokenization and Packing Interface

Tokenization and Packing Interface gives the conceptual and mathematical layer for full dataset assembly. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

5.1 Token-count budgets

Token-count budgets is part of the canonical scope of full dataset assembly. 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 source set, 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 Sequence packing

Sequence packing is part of the canonical scope of full dataset assembly. 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 mixture weight, 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 Document-boundary masks

Document-boundary masks is part of the canonical scope of full dataset assembly. 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 token budget, 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 Padding/truncation

Padding/truncation is part of the canonical scope of full dataset assembly. 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 manifest, 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 Packed-shard statistics

Packed-shard statistics is part of the canonical scope of full dataset assembly. 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 shard, 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. Final Verification

Final Verification gives the conceptual and mathematical layer for full dataset assembly. The local variables in this section should be read as pipeline objects: documents, records, tokens, filters, weights, shards, and manifests.

6.1 Shard count

Shard count is part of the canonical scope of full dataset assembly. 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 source set, 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 Token count

Token count is part of the canonical scope of full dataset assembly. 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 mixture weight, 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 Source proportions

Source proportions is part of the canonical scope of full dataset assembly. 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 token budget, 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 rebuild

Reproducible rebuild is part of the canonical scope of full dataset assembly. 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 manifest, 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 Smoke-test data loader

Smoke-test data loader is part of the canonical scope of full dataset assembly. 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 shard, 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 source set example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
2. (*) Build a synthetic mixture weight example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
3. (*) Build a synthetic token budget example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
4. (**) Build a synthetic manifest example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
5. (**) Build a synthetic shard example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
6. (**) Build a synthetic split example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
7. (**) Build a synthetic packing example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
8. (***) Build a synthetic source set example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
9. (***) Build a synthetic mixture weight example, compute its validation signal, and explain which downstream stage would fail if the signal were wrong.
10. (***) Build a synthetic token budget 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 [Contamination and Dedup Audits](../05-Contamination-and-Dedup- Audits/notes.md). 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