| # Scaled replication: Automatic Layer Selection for Hallucination Detection |
|
|
| This bundle implements a pinned, scaled CoQA replication for |
| `mistralai/Mistral-7B-Instruct-v0.3` and |
| `meta-llama/Llama-3.1-8B-Instruct`. It is deliberately split into: |
|
|
| 1. a capped A100 GPU stage for stochastic answer generation, the released |
| Ministral judge procedure, Llama hidden states, and a 64-record CUDA |
| Mistral cross-check; |
| 2. memory-bounded Mistral hidden-state extraction on one TPU v5e chip; |
| 3. CPU TwoNN, reconstructed FEPoID selection, and per-layer MLP probes. |
|
|
| The experiment uses 1,000 flattened CoQA train examples and 1,000 validation |
| examples per model. The paper uses 10,000 train/validation examples and a |
| 7,983-example CoQA test set, then averages five QA datasets. Ten temperature-1 |
| candidate answers are omitted because FEPoID does not consume them. |
|
|
| ## Critical qualification |
|
|
| **An exact reproduction is virtually impossible from the released artifacts.** |
| The authors did not release their sampled `best_answer` records, judge labels, |
| model/dataset revisions, environment lock, selected layers, or probe outputs; |
| their `ID_twonn.py` also stops after producing the ID curve and does not |
| implement FEPoID. This bundle is an independent replication with pinned current |
| snapshots and documented reconstruction choices—not evidence that the exact |
| published AUROCs were regenerated. |
|
|
| See `configs/scaled_coqa.json` for every pinned revision and parameter. |
|
|
