citation-extractor-fast — training data and evaluation

The exact training cohort and frozen evaluation for citation-extractor-fast: legal source text paired with citation quotes, kinds and Unicode source offsets. This release includes the training recipe, checkpoint-bound code, split manifests, original evaluation predictions and a score-reproduction check.

Split	Documents	Contents
train	8,015	Actual examples used for the published one-epoch fine-tune
test	200	Frozen federal holdout documents and silver citation references

The original 8,129-candidate training archive and the 114 fit exclusions are included under reproduction/. The admitted training cohort is identified by the original common-fit manifest and final checkpoint's ordered document IDs. Train and test are disjoint by document ID, source family and exact source text.

Load the data

from datasets import load_dataset

data = load_dataset("thurstonpowers/citation-extractor-fast")
example = data["train"][0]
print(example["text"])
print(example["citations"])
print(example["messages"])

Each row contains:

• document_id, document_type, source_family: source identity and grouping.
• raw_sha256: SHA-256 of the exact UTF-8 source text.
• size_band, complexity_band: original sampling strata.
• text: complete original user-input text, preserving characters and line endings.
• citations: objects with quote, start, end, kind.
• messages: the actual system/user/assistant training messages; test rows contain system/user messages, with reference answers in citations.

Offsets are zero-based Unicode codepoints, with exclusive end. Every citation satisfies text[start:end] == quote. Repeated occurrences have separate offsets. The original assistant strings and prompt bytes are preserved. SYSTEM_PROMPT.txt SHA-256: 8e2451afcf15498b0aab55b490fd0536342fcf097ce5dc3abd99958556de4f85.

Label provenance

Sources are U.S. federal statutes, regulations, bills, acts and proposed rules. Citation labels were assembled from three teacher draws per source, reconciled using cross-draw support and exact source alignment. Repeated literal quote occurrences were projected to their source positions. These are silver annotations: coverage and teacher interpretation can contain errors.

The release preserves the source IDs/text hashes, original training archive, reference answers, and reconciliation implementation. Independent legal adjudication remains a separate step. See provenance/data-release.json for artifact identities and label-generation details.

Reproduce the published evaluation

The original eval200 includes 200 predictions per arm for untrained Nemotron, GPT‑6.1 Sol and fine-tuned Nemotron. Raw generated text, failure status, caps, timeouts, token counts, durations and scoring-relevant usage values are included. All original failures remain in the denominator.

Document type	Documents	Untrained Nemotron F1	GPT‑6.1 Sol F1	Fine-tuned original F1	Fine-tuned + timeout retries F1
Acts	37	3.32%	23.21%	55.76%	61.88%
Bills	44	0.73%	13.19%	49.01%	49.01%
Proposed rules	37	0.88%	31.43%	32.58%	32.58%
Regulations	44	0.59%	30.01%	38.08%	34.15%
Statutes	38	1.33%	17.57%	43.64%	43.64%
Overall	200	1.31%	23.18%	44.23%	44.55%

These are exact-quote multiset F1 scores. evaluation/comparison.json includes all metrics and strata; evaluation/comparison-by-document-type.md includes precision, recall and failures. The original fine-tuned arm has 31 caps and 19 600-second timeouts. The larger-budget follow-up below is recorded separately.

Completed longer-budget supplement

supplements/longer-budget-20261105/ preserves both retry stages, their fixed assignments, raw outputs, protocols, completion/capacity receipts and all scores. The first retry raised the deadline to 1,800 seconds. Four EOS outputs were retained, then the remaining 15 cases were run with 32,768 output tokens, 65,536 context tokens and the same deadline. The final 19-case cohort contains five complete JSON arrays and 14 capped outputs.

The last table column retains 181 original predictions and substitutes these 19 fixed retry outcomes. All 200 documents remain in the denominator. Overall quote precision/recall/F1 become 65.07% / 33.87% / 44.55%, with 45 caps and zero timeouts across the combined 200 outcomes. The original result files remain unchanged. The verifier recomputes every original and supplemental score group and checks that the other 181 prediction records are unchanged.

The two retry stages add 30 saved predictions and one caller-interrupted attempt to the original 200. The supplement accounts for 231 started attempts, including the interrupted token trace, and records cumulative time/tokens for the 230 saved prediction records. Thinking-disabled decoding was used throughout the fine-tuned runs; the larger allowance funds generated citation JSON. Repeated citations and drifting offsets remained substantial failure modes.

Download an immutable snapshot and verify it using Python's standard library:

from huggingface_hub import HfApi, snapshot_download

repo = "thurstonpowers/citation-extractor-fast"
revision = HfApi().dataset_info(repo).sha
snapshot = snapshot_download(repo, repo_type="dataset", revision=revision)
print(snapshot)

python /path/to/snapshot/verify_release.py /path/to/snapshot

This checks every file checksum, original training archive and code hashes, ordered cohort membership, exact source spans, split separation and every published score group across all three arms. It reports results without a GPU or API calls. Score recomputation uses preserved generated outputs; fresh model inference is available through the included frozen workers and protocols.

Reproduce training

Base model: nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16, immutable revision dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f.

Recipe: full language-weight fine-tuning; native BF16 parameters with architecture-specific dtype exceptions; FP32 Adafactor state; one epoch; microbatch one, accumulation four; LR 5e-6; 101 warmup updates; cosine schedule to 0.1 of initial LR; seed 102301; 2,004 updates; complete sequences up to 32,768 tokens. Exact arguments and tensor/data contracts are in reproduction/.

1. Install a compatible CUDA/PyTorch environment using the versions recorded in reproduction/training-environment.json. Training was run on NVIDIA GB10 with Python 3.12, PyTorch 2.13.0+cu130 and Transformers 5.15.1. The complete package inventory and original image digest are included.

2. Download the pinned base model into the Hugging Face cache:

   hf download nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 --revision dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f
   

3. Provide fresh GPU thermal telemetry at thermal/status.json, with ISO-8601 UTC timestamp and numeric max_temperature_c. The included thermal helper uses pause at 87°C and resume below 80°C. Keep your host safety guard active.

4. From the dataset snapshot directory, run the actual frozen training pipeline:

   CUBLAS_WORKSPACE_CONFIG=:4096:8 python code/training/run_citation_student_pipeline.py \
     --data reproduction/train-candidates.jsonl.gz \
     --common-fit reproduction/common-fit.json \
     --model-key nemotron --output reproduced-run \
     --thermal-status thermal/status.json \
     --deterministic --allow-nonbitwise-trajectory
   

The pipeline qualifies optimizer updates and state restoration, then trains the admitted cohort. reproduction/encoded-manifest.jsonl records every original token sequence fingerprint for tokenization checks. Final checkpoint code hashes match the supplied training helpers byte-for-byte.

Numerical reproducibility: the recorded qualification restored parameters, buffers, optimizer, RNG, input encoding and data position exactly. Subsequent GPU-kernel trajectories varied in matched runs; the measured evidence and qualification receipt are included. The published model's checkpoint/export/ reload equality is separately documented. The data, recipe, token fingerprints and evaluation scores provide explicit checks for replication.

License and source attribution

Federal source text comes from U.S. government publications identified by the preserved source IDs. The dataset compilation, annotations and reproduction utilities are released under Apache-2.0, following the project's dataset-release convention. See LICENSE. The model weights retain their own license in the linked model repository. API response envelopes are reduced to scoring-relevant usage fields; complete generated answers and failure outcomes are preserved.