The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Olmo 3 DPO calibration study: raw outputs
Raw per-question and per-response outputs and run logs from a study of why DPO makes Olmo 3 7B Instruct
overconfident on multiple-choice questions. The code, summaries, pre-registered predictions and write-up ("Why Does DPO Make Olmo 3 Overconfident?") live in the GitHub repo
edwin-huang/olmo3-dpo-calibration; this repository holds the files too large to keep there. Paths match the repo's olmo-calib/ folder,
so a file referenced there as results_ckpt/<run>/... is at the same path here.
What is here
| Files | Contents |
|---|---|
**/*questions.jsonl, **/letters.jsonl |
Per-question A-D letter probabilities (MMLU, ARC; 5-shot letters, chat format, self-consistency) used for T*, ECE and AUROC |
**/*responses.jsonl |
Model generations: 300 held-out chat prompts, IFEval, BigBenchHard (Ai2 protocol and our 3-shot harness) |
**/*cloze_rows.jsonl |
Per-question cloze scores |
**/*.log, results_box2/, results_q25/runs/, results_smoke/ |
Training and evaluation logs |
Trained models are in ehuang5/olmo3-calibration-models.
Sources and license
Outputs come from Ai2's Olmo 3 7B checkpoints (Apache-2.0) and from DPO models trained on Ai2's Dolci preference data (allenai/Dolci-Instruct-DPO, allenai/Dolci-Think-DPO-7B, both ODC-BY). Prompts and questions come from those datasets and from public benchmarks (MMLU, ARC-Challenge, IFEval, BIG-Bench Hard). This repository is released under ODC-BY; please attribute Ai2's Olmo 3 and Dolci releases as well as the benchmarks.
- Downloads last month
- 816