Hugging Face's logo

Datasets:
GwendalTsang
/
automatic-layer-selection-reproduction-bundle

Dataset card Files
xet
 Community
automatic-layer-selection-reproduction-bundle / HANDOFF.md
GwendalTsang's picture
GwendalTsang
Add placeholder for full TPU v5e-8 Mistral bucket
52db592 verified
preview code
|
Raw
History Blame Contribute Delete
10.7 kB

Handoff: Automatic Layer Selection for Hallucination Detection

Last updated: 2026-07-20 17:06 UTC

Paper: Automatic Layer Selection for Hallucination Detection
arXiv: 2605.26366
OpenReview: srAuLCB3Ca

Read this first

An exact reproduction of the paper is virtually impossible from the released artifacts. The authors did not release the generated best_answer records, the ten uncertainty candidates per input, judge decisions/labels, exact model and dataset snapshots, dependency lock, selected layers, trained probe states, or raw results. The released method/ID_twonn.py computes a TwoNN curve but does not implement the paper's FEPoID peak selector. In addition, the paper says "exact string match" while the released labeling code uses substring containment. Any result produced here is therefore a pinned independent, scaled replication, not a regeneration of the authors' exact AUROCs.

The current experimental scope is 1,000 flattened CoQA training examples and 1,000 flattened validation examples for Llama. The paper uses 10,000 training examples, the full 7,983-example CoQA test set, and averages five QA datasets. Do not describe the current run as verification of the paper's five-dataset Table 2 mean.

Persistent assets

Current live Llama run

Job ID: GwendalTsang/6a5e54a9d09dc1f57c6bd8f8
Job name: fepoid-llama-coqa-2k-v3
Hardware: one A100 80 GB (a100-large, $2.50/hour)
Hard timeout: 110 minutes (maximum substantive-run charge about $4.59)
Docker image: pytorch/pytorch:2.7.1-cuda12.8-cudnn9-runtime

At the snapshot time, the job was RUNNING. Both Llama answer-generation splits had completed: 1,000 train + 1,000 validation examples in 162.49 seconds. The job was downloading/loading the pinned Ministral judge. It is remote and will continue after the current TPU session is terminated.

Monitor it with:

hf jobs inspect GwendalTsang/6a5e54a9d09dc1f57c6bd8f8 --format json
hf jobs logs GwendalTsang/6a5e54a9d09dc1f57c6bd8f8 --tail 200
hf jobs wait GwendalTsang/6a5e54a9d09dc1f57c6bd8f8 --timeout 30s --format json

The first two launch attempts are useful negative provenance, not successful experiments:

  • 6a5e5423d09dc1f57c6bd8f0: inline bash -lc was parsed incorrectly; exit 127 before inference.
  • 6a5e545dd09dc1f57c6bd8f4: launcher worked but scikit-dimension had an undeclared Matplotlib import; failed before model loading. The pinned dependency was added before v3.

The Jobs canary succeeded: https://huggingface.co/jobs/GwendalTsang/6a5e5324d09dc1f57c6bd897.

Pinned inputs

  • Model: meta-llama/Llama-3.1-8B-Instruct
  • Model revision: 0e9e39f249a16976918f6564b8830bc894c89659
  • Judge: mistralai/Ministral-8B-Instruct-2410
  • Judge revision: 2f494a194c5b980dfb9772cb92d26cbb671fce5a
  • Dataset: stanfordnlp/coqa
  • Dataset revision: 0d9e9952f1ef6e5415492d3d84b5873259137e3c
  • Generation: seed 2024, temperature 0.1, 30-token cap
  • FEPoID forward horizon: w=7
  • The ten temperature-1 uncertainty candidates are intentionally omitted because FEPoID itself does not consume them.

Expected persisted output

Bucket prefix: coqa-2k-llama-20260720/

  • environment.json
  • prepared/*.jsonl: generated answers, FST answers, labels, judge outputs, model/dataset revisions, and source indices
  • hidden_states/llama/{full,fst}/{train,validation}/manifest.json
  • 16 SafeTensors shards in total: four views/splits times four 250-record shards
  • Every shard contains:
    • hidden_states: BF16 [records, 32, 4096]
    • labels: int64 [records]
    • record_indices: int64 [records]
  • metrics/llama_{full,fst}_layers.csv
  • metrics/llama_{full,fst}_summary.json
  • gpu_stage_summary.json
  • SUCCESS

Layer semantics match Transformers output.hidden_states[1:]: layers 0–30 are post-block/pre-final-RMSNorm representations and layer 31 is post-final-block, post-final-RMSNorm. Inputs are left padded; index -1 is the last non-padding token. use_cache=False is used for extraction. Full and FST answers are retokenized using the paper release's prompt-plus-answer construction.

Verify the Llama artifact after completion

Do not treat COMPLETED alone as proof. Check the bucket and the success marker:

hf buckets list GwendalTsang/fepoid-llama-hidden-states \
  --recursive --human-readable --format json

mkdir -p outputs/llama-coqa-2k
hf buckets sync \
  hf://buckets/GwendalTsang/fepoid-llama-hidden-states/coqa-2k-llama-20260720 \
  outputs/llama-coqa-2k

test -f outputs/llama-coqa-2k/SUCCESS
find outputs/llama-coqa-2k/hidden_states -name '*.safetensors' | wc -l

The shard count should be 16. Validate every manifest's record count, shape, dtype, and SHA-256 against the downloaded shard. Also ensure every prepared split has 1,000 rows and both labels are present in train and validation before trusting AUROC. If the job fails after exporting some shards, do not mix partial and rerun output under the same prefix; change the prefix in scripts/run_llama_hf_job.sh and relaunch.

Local workspace layout

  • /content/main.tex, /content/sections/, /content/tables/: extracted paper source used for the audit.
  • /content/paper-code/: authors' repository at commit def3cb6.
  • /content/repro_automatic-layer-selection-for-hallucination-detection/: scaled independent reproduction bundle.
  • /content/.trackio/logbook/: canonical five-claim Trackio logbook scaffold.
  • /content/mistral-tpu8-hidden-states/: published eight-device TPU extractor, clean at commit 10cc151 on main.

Important reproduction files:

  • configs/scaled_coqa.json: pinned revisions, scale, and limitations.
  • scripts/hybrid_gpu_stage.py: Llama generation, judging, BF16 extraction, sharding, TwoNN/FEPoID, and MLP probes. It now supports --models llama without loading Mistral.
  • scripts/run_llama_hf_job.sh: exact dependency lock and v3 Job command.
  • scripts/repro_common.py: reconstructed FEPoID, FST scanner, TwoNN, probes.
  • scripts/analyze_mistral_tpu.py: consumes provided TPU Mistral artifacts.

Local checks completed:

  • Python compilation passed for all three Python scripts.
  • hybrid_gpu_stage.py --self-test passed.
  • A SafeTensors round trip confirmed BF16 [N,32,4096] export and sharding.
  • The Mistral TPU repository has 5 passing unit tests and a successful GitHub Actions run: https://github.com/GwenTsang/mistral-tpu8-hidden-states/actions/runs/29761410557.
  • A real one-core TPU v5e Mistral smoke test passed. Eight-device orchestration was not available in the original session, so the extractor enforces --expected-world-size 8 to prevent silent underuse.

Mistral handoff contract

The user explicitly said they will provide the Mistral hidden states later. Do not spend GPU/TPU money regenerating Mistral unless the user changes that instruction. Ask for or locate the provided Mistral output, then validate it.

Preferred Mistral shape is BF16 hidden_states=[N,32,4096] plus JSONL metadata with stable record order and labels for full/FST views. The published extractor also saves embedding=[N,4096]; the analysis does not require the embedding. Run its validator first, then adapt/execute:

python /content/mistral-tpu8-hidden-states/validate_output.py /path/to/mistral-output
python scripts/analyze_mistral_tpu.py --help

Never compare Llama and Mistral AUROC unless they use the same flattened CoQA indices, split definitions, prompt format, generated-answer policy, correctness labeling policy, FST rule, probe split, and layer indexing.

Remaining work, in order

  1. Monitor v3 to a terminal state and perform the artifact audit above.
  2. Download only small summaries/manifests first; record wall time, estimated cost, label balance, selected layers, full/FST AUROCs, last-layer AUROCs, and oracle AUROCs in the logbook.
  3. Add a validator for all Llama manifests/shards if the current Job succeeds.
  4. When the user supplies Mistral activations, validate and analyze them with the same split/probe conventions. Do not regenerate Mistral by default.
  5. Populate Trackio claim pages:
    • Claim 1: Llama layer-wise AUROC curve and whether its maximum is intermediate.
    • Claim 2: scaled CoQA FEPoID versus last/max-ID/oracle; clearly state that five-dataset means 0.7253/0.8531 were not reproduced.
    • Claim 3: not empirically tested yet; a summarization run is still needed for substantive verification.
    • Claim 4: time the reconstructed selector separately from activation extraction and probes; the paper's 10.14-second cross-benchmark mean cannot be regenerated from author artifacts.
    • Claim 5: compare the same records under full versus FST and report paired AUROC changes.
  6. Build the required Chenruishuo/posterly poster with --strict-polish, add poster_embed.html to Executive summary, and pin it below the summary.
  7. Register the dedicated reproduction folder as the Conclusion artifact.
  8. Validate and publish:
curl -sL https://huggingface.co/spaces/ICML-2026-agent-repro/challenge/raw/main/scripts/validate_icml_logbook.py | \
  python3 - --space GwendalTsang/repro-automatic-layer-selection-for-hallucination-detection

trackio logbook publish \
  GwendalTsang/repro-automatic-layer-selection-for-hallucination-detection

The logbook is only a scaffold at handoff; it has not been validated or published. Its Executive summary must prominently repeat that exact reproduction is virtually impossible and distinguish the scaled independent replication from the paper's full experiment.