---
pretty_name: Mistral-7B TPU hidden-state extractor verification (CoQA references)
language:
- en
task_categories:
- feature-extraction
tags:
- mistral
- hidden-states
- tpu
- pytorch-xla
- coqa
---

# Mistral-7B TPU hidden-state extractor verification

> **Verification-only artifact — not a reproduction of the paper's AUROC.**

This dataset contains last-token embeddings and hidden states extracted from
1,000 flattened CoQA validation question/reference-answer pairs with
`mistralai/Mistral-7B-Instruct-v0.3`. It verifies that the memory-bounded TPU
v5e-1 extraction path runs successfully. It does **not** contain the Mistral
`best_answer` generations or hallucination labels required to reproduce Table
2 of *Automatic Layer Selection for Hallucination Detection*.

The paper authors' public repository does not contain its `prepared_data`
records. Do not treat this artifact as evidence for or against the paper's
FEPoID or hidden-state-probing AUROC claims.

## Tensor layout

Each `states/validation/shard-*.safetensors` file contains BF16 tensors:

| Key | Shape | Meaning |
| --- | --- | --- |
| `embedding` | `[N, 4096]` | Last-position input-token embedding |
| `hidden_states` | `[N, 32, 4096]` | Last-position state for every Mistral layer |

Layers 0–30 are post-transformer-block states. Layer 31 is post-final-RMSNorm,
matching the tensors selected as `outputs.hidden_states[1:]` by the released
CUDA code. `metadata-validation.jsonl` maps each source index to a shard and
offset. `inputs-validation.jsonl` contains the exact prompt-answer text and its
SHA-256 digest.

## Memory-bounded TPU method

- Loads `MistralModel`, omitting the unused language-model head.
- Uses BF16, `use_cache=False`, `XLA_NO_SPECIAL_SCALARS=1`, and PyTorch/XLA JIT.
- Never enables `output_hidden_states=True`.
- Hooks only `[:, -1, :]` after each layer and performs one host transfer per batch.
- Uses static sequence buckets of 128, 256, 512, 1,024, and 2,048 tokens.
- Saves resumable SafeTensors shards rather than accumulating the dataset in RAM.

The 1,000-record extraction completed in 215.8 seconds after model placement,
using the 256-, 512-, and 1,024-token buckets without truncating any record.

## Loading

```python
import json
from huggingface_hub import hf_hub_download
from safetensors import safe_open

repo_id = "GwendalTsang/mistral-7b-hidden-states-tpu-verification"
metadata_path = hf_hub_download(
    repo_id, "metadata-validation.jsonl", repo_type="dataset"
)
row = json.loads(open(metadata_path, encoding="utf-8").readline())
shard_path = hf_hub_download(repo_id, row["shard"], repo_type="dataset")
with safe_open(shard_path, framework="pt", device="cpu") as shard:
    embedding = shard.get_tensor("embedding")[row["offset"]]
    hidden_states = shard.get_tensor("hidden_states")[row["offset"]]
```

See [`scripts/extract_mistral_hidden_states_tpu.py`](scripts/extract_mistral_hidden_states_tpu.py)
for reproduction and for processing paper-compatible JSONL records containing
`context`, `question`, `best_answer`, and `label`. The extractor supports both
`--answer-view full` and the authors' rule-based
`--answer-view first_sentence`.

## Provenance

- Model revision: `c170c708c41dac9275d15a8fff4eca08d52bab71`
- CoQA revision: `0d9e9952f1ef6e5415492d3d84b5873259137e3c`
- Paper code revision: `def3cb6d262c11d252e6c5a6b7e04375b94b54db`
- Paper: https://arxiv.org/abs/2605.26366
- Released code: https://github.com/DesoloYw/Automatic-Layer-Selection-for-Hallucination-Detection