Search is not available for this dataset
n int64 468 513 | wrong_rate float64 0.3 0.59 | auroc_max_span_risk float64 0.81 0.88 | threshold float64 0.97 0.99 | flag_rate float64 0.09 0.09 | precision_at_threshold float64 0.89 0.98 | recall_at_threshold float64 0.14 0.27 |
|---|---|---|---|---|---|---|
513 | 0.592593 | 0.814105 | 0.984372 | 0.091618 | 0.893617 | 0.138158 |
470 | 0.404255 | 0.882124 | 0.967067 | 0.093617 | 0.977273 | 0.226316 |
468 | 0.299145 | 0.861215 | 0.985917 | 0.089744 | 0.904762 | 0.271429 |
llm_clearest_text probes and training data
Linear probes, and the data they were trained on, for github.com/rrenaud/llm_clearest_text. That project marks up chat answers with probes on the model's own hidden states. Result: the probes are not accurate enough to be useful in long-form chat (see the repo README).
There is one folder per model: Qwen2.5-1.5B-Instruct, Qwen2.5-7B-Instruct, Qwen3.6-27B (thinking off).
| File | Contents |
|---|---|
trivia.jsonl |
8,000 TriviaQA (rc.nocontext, validation) questions with the model's greedy short answer, correct (alias match), and the logprob baselines lp_first, lp_mean, lp_min |
responses.jsonl |
about 600 chat responses to Dolly-15k prompts plus 30 advice prompts |
style_labels.jsonl |
the same responses, with every sentence labeled boilerplate, filler or substantive by Claude |
probe_correctness.npz/.json |
logistic-regression probe on the last answer token's hidden state, with layer and metrics |
probe_boilerplate.npz, probe_filler.npz, probe_style.json |
sentence probes on mean-pooled hidden states, with thresholds and metrics |
eval_prose_risk.json |
transfer to one-sentence answers: AUROC, plus precision and recall at the UI threshold |
The hidden-state caches are not included. Regenerate them with scripts/train_probes.sh in the repo.
TriviaQA is Apache-2.0. Dolly-15k is CC BY-SA 3.0.
- Downloads last month
- -