REIM SimulatedEstimator from psim (Petri + ConfirmedDetoxify)
Gated. Queries are drawn from the description-query psim
(meta-llama/Meta-Llama-3-8B + singhalrk/rare-input-model-description-query-checkpoint-5964), not from WildChat. n is a bit under 1M
because invalid generated queries were dropped.
from datasets import load_dataset
ds = load_dataset("rangell/reim-sim-psim-petri-detoxify", "qwen3-8b", split="train")
| Config | Target | n | Sampling |
|---|---|---|---|
qwen3-8b |
Qwen/Qwen3-8B |
999,950 | 50 × 20,000 |
deepseek-llama |
deepseek-ai/DeepSeek-R1-Distill-Llama-8B |
960,150 | 64 × 15,625 |
nemotron |
nvidia/Nemotron-Cascade-14B-Thinking |
960,193 | 64 × 15,625 |
Same judge stack as the MC companion
rangell/reim-mc-wildchat-petri-detoxify:
Petri (gpt-oss-120b) + ConfirmedDetoxify on the same responses. Unweighted
simulated rates are in {config}/summary.json.
Qwen3-8B was Petri-then-backfill (scalar simulated_rates in that summary use
the backfill reward dim). DeepSeek-Llama and Nemotron used the composite judge
in the array.
- Downloads last month
- 10