ShoaibSSM/hallucination-portability-revision

Hallucination probe portability revision

First GPU round is ready: new representation corpora for Ministral-3-8B-Base-2512 and Qwen3-8B, using pinned model/dataset revisions. Each has 5,000 training, 1,000 development, and 1,000 test SQuAD v2 examples. Development/test contexts do not overlap. This is a new controlled training/adaptation dataset, not a claim to reproduce the old split.

Run in your independent terminal:

cd /home/jovyan/hallucination-portability
export HF_HOME="$PWD/artifacts/hf_cache"
source ~/.bashrc
bash scripts/launch_stage1.sh

The launcher prints the PID and log path. It uses nohup; this session can close. Monitor:

cd /home/jovyan/hallucination-portability
tail -f "$(cat logs/stage1_latest_log.txt)"

Continuous Hugging Face sync is enabled by default. It creates/uses a private dataset repository ShoaibSSM/hallucination-portability-revision, uploading completed shards, scripts/configs, logs/results, and later probe checkpoints under artifacts/probes. It excludes HF caches, credentials, original files and incomplete shards. Single upload worker, durable retry state, 30-second polling, tensor payloads before completion markers. Uploads continue draining after GPU work ends; monitor logs/hf_sync.log. HF authentication was verified after sourcing ~/.bashrc. Actual remote artifact upload remains unverified. Run the login command in your terminal; never paste the token into chat/source. Set HF_SYNC_REPO=your-namespace/your-repo before launching to change the destination. If you deliberately want local-only work, HF_SYNC=0 bash scripts/launch_stage1.sh; later HF_HOME="$PWD/artifacts/hf_cache" .venv/bin/python scripts/hf_sync.py uploads the backlog.

Future best checkpoints should have immutable versioned filenames so every improvement survives the polling interval; the sync scans those plus final results. Live upload success is independent from GPU queue exit status.

After an interruption, use the same launch command. Committed shards are checksum-verified and skipped; incomplete shard payloads are regenerated. A project GPU lock prevents two copies of this queue running together. Exit status is written to logs/stage1_queue.exit (0 means success); the newest log is authoritative while a job is running.

The queue performs a 100-row pilot and integrity validation for each model before continuing to 7,000 rows. It benchmarks batch sizes 1/2/4/8/16/32 on fixed examples and uses the fastest candidate under its measured 32 GiB allocated-memory envelope. OOM causes batch-size reduction without silently skipping examples. No unrelated processes are stopped.

Outputs live under artifacts/stage1_ministral3_8b/ and artifacts/stage1_qwen3_8b/: full example manifests, generated text/token IDs, FP16 hidden-state mmap shards, per-token log-probability/entropy, run configs, batch benchmark, progress, completion and validation JSON. Storage budget: approximately 359 GiB tensors total, plus roughly 30 GiB model downloads/cache and metadata. Runtime depends on lengths and tuned batch size; the 10-example real Mistral write test took ~7 seconds after loading, which is too small to promise total runtime.

Qwen3 has 36 layers, Ministral 3 has 34. The pipeline preserves these dimensions. No silent crop/reshape. Both models use BF16 compute with FP16 storage; these are separately recorded new-model protocols. Plain context/question/answer prompts intentionally match the legacy prompt family; no native chat/thinking template is silently added.

This round does not train probes or produce factuality labels. judge_representations.py is a draft for the next stage and is deliberately not in the launch queue yet. Next: validated multi-judge labels, then target-trained controls, frozen transfer and source-initialized probe fine-tuning under matched target-label budgets. Human annotations remain human work.

Validation already performed: real pretrained Mistral extraction for 10 rows, checksum/padding/ID verification, resume without regeneration, small-model GPU single/batch comparison, and storage/metric/extraction tests. Qwen's architecture path is tested with tiny random weights; the queue's 100-row pilot is its first real-checkpoint validation.

Original material remains unchanged in /home/jovyan/Previous-files. See docs/PROJECT_AUDIT.md, docs/REVIEWER_REQUIREMENTS_MATRIX.md, docs/EXPERIMENT_PLAN.md, docs/PIPELINE_PERFORMANCE_AUDIT.md, and docs/RESOURCE_PLAN.md. Missing transfer source/Asta/LaTeX and checkpoint provenance discrepancies remain explicitly open. Literature verification is incomplete.

Saved-score reproduction (CPU only):

bash scripts/reproduce_saved_scores.sh --output results/reproduction/legacy_point_metrics
# Optional longer CPU uncertainty analysis, independent of GPU acquisition:
nohup bash scripts/reproduce_saved_scores.sh --bootstrap 2000 --output results/reproduction/legacy_bootstrap > logs/legacy_bootstrap.log 2>&1 < /dev/null &

Environment is project-local .venv; dependencies are recorded in requirements-lock.txt. Do not run legacy upload helpers: source-generation scripts contain an automatic-upload path and embedded token defaults; the revised tools do not publish anything.

Model replacement: the active queue now uses Ministral 3 8B Base (December 2025), not legacy Mistral-v0.3. Official model card: https://huggingface.co/mistralai/Ministral-3-8B-Base-2512 . The base model matches the existing completion-style prompt and supports later controlled post-training. The old Mistral real smoke test remains historical; new Ministral has tokenizer/config and tiny-architecture checks, followed by a mandatory real 100-row runtime pilot.

Stage 1 complete; next: three-judge labeling

Both modern-model corpora have 7,000 validated local rows. See completion audit for timing, truncation caveats and backup status. Full hidden tensors remain local; HF sync now defaults to essential metadata, labels, results, logs and probe checkpoints. The revision HF repository is public with the user's authorization.

Stage 2 uses saved responses and pinned Mistral-v0.2, Qwen3-8B and Phi-4-mini judges. The old Mistral is only the continuity judge, not a new generation/training corpus. See stage 2 runbook.

bash scripts/launch_stage2.sh
tail -f "$(cat logs/stage2_latest_log.txt)"

Each full labeling run has a 100-row pilot gate. Invalid judge responses remain null. Review label coverage and agreement before training. Full runs are user-launched.