lucasfrag/fact-checking-nei-saes-data

Fact-Checking NEI SAEs — data

Companion data for lucasfrag/fact-checking-nei-saes.

Per base model (llama-3.1-8b-instruct/, qwen3-8b/, gemma-3-4b-it/, deepseek-r1-distill-llama-8b/)

file	content
examples.jsonl	the exact ordered list of training claims (VitaminC + FEVER train splits, gold evidence), one JSON per line: claim, evidence (list), label (SUP/REF/NEI), source. Line i is prompt i of the harvest, so activations can be regenerated deterministically.
decisions.safetensors	residual-stream activations at the decision position (last prompt token) for every claim, in the same order: acts [N, d_model] fp16 (bf16 for Gemma: one residual dimension of Gemma-3 is ~32,768 on every token and overflows fp16), label (0 SUP, 1 REF, 2 NEI), source (0 VitaminC, 1 FEVER). Enough to reproduce the NEI-feature analysis without re-running the base model.

The full per-token training activations (~1 TB per model) are not released; they are regenerable from examples.jsonl with the released code.

SAE viewer data (viewer/<model>/layer-<L>/)

Data for the SAE viewer: index.json (per-feature summary) and atoms/<id // 1000>/<id>.json (one file per feature: top / random claims at the verdict position with their evidence, activation histograms by gold label, nearest features by decoder cosine, logit-lens effects). Only features that fire at the verdict position are listed in index.json.

steering/<id>.json (only for steered features, e.g. 19219): dose-response curves and example generations of the steering experiment (adding α × the feature's decoder direction while the model writes its verdict on 500 AVeriTeC dev claims). These files contain AVeriTeC claims and evidence excerpts and are therefore distributed under CC BY-NC 4.0, the license of AVeriTeC (Schlichtkrull et al., 2023, arXiv:2305.13117), not under the CC BY-SA 3.0 license of the rest of this dataset.

License and attribution

This dataset is released under CC BY-SA 3.0, the license of its sources (share-alike). The claim and evidence texts in examples.jsonl come from:

• VitaminC — Schuster, Fisch & Barzilay (2021), tals/vitaminc, CC BY-SA 3.0;
• FEVER (gold-evidence version copenlu/fever_gold_evidence) — Thorne et al. (2018), CC BY-SA 3.0;

both derived from Wikipedia. The activations in decisions.safetensors were computed from these texts with the base models listed in the companion model repository (see their licenses there).

Derived from VitaminC (Schuster et al., 2021) and FEVER (Thorne et al., 2018); see their licenses.