REIM Monte Carlo on WildChat (Petri + ConfirmedDetoxify)
Gated. Source queries are from
singhalrk/wildchat_filtered_english_description_label_no_moderation
(filtered English WildChat). Join back to that set with conversation_hash.
One config per target LM. k=1 response per query. Petri (gpt-oss-120b, three
unprompted traits in one call) and ConfirmedDetoxify (Detoxify heads + Llama-3.1-8B
confirmation) on the same responses.
from datasets import load_dataset
ds = load_dataset("rangell/reim-mc-wildchat-petri-detoxify", "qwen3-8b", split="train")
| Config | Target | n |
|---|---|---|
qwen3-8b |
Qwen/Qwen3-8B |
1,012,880 |
deepseek-llama |
deepseek-ai/DeepSeek-R1-Distill-Llama-8B |
1,012,880 |
nemotron |
nvidia/Nemotron-Cascade-14B-Thinking |
1,012,880 |
Thresholded rates (P(score ≥ 0.1 / 0.5 / 0.9)) live in
{config}/summary.json, not in the parquet. {config}/config.json is the
run config.
Qwen3-8B was judged with Petri first, then ConfirmedDetoxify was backfilled
onto those responses. DeepSeek-Llama and Nemotron used
CompositeJudge([Petri, ConfirmedDetoxify]) in the sampling array, so they
also have per-trait Petri *_raw / summary_* / justification_* columns
that Qwen lacks.
- Downloads last month
- 10