Datasets:
query_id int64 0 499 | rollout_id int64 0 127 | phi_conf float64 0 0.96 |
|---|---|---|
0 | 0 | 0.064639 |
0 | 1 | 0.064639 |
0 | 2 | 0.066072 |
0 | 3 | 0.066072 |
0 | 4 | 0.034352 |
0 | 5 | 0.066072 |
0 | 6 | 0.066072 |
0 | 7 | 0.064639 |
0 | 8 | 0.066072 |
0 | 9 | 0.064639 |
0 | 10 | 0.034352 |
0 | 11 | 0.064639 |
0 | 12 | 0.066072 |
0 | 13 | 0.066072 |
0 | 14 | 0.066072 |
0 | 15 | 0.066072 |
0 | 16 | 0.034352 |
0 | 17 | 0.066072 |
0 | 18 | 0.064639 |
0 | 19 | 0.064639 |
0 | 20 | 0.064639 |
0 | 21 | 0.064639 |
0 | 22 | 0.061399 |
0 | 23 | 0.066072 |
0 | 24 | 0.064639 |
0 | 25 | 0.064639 |
0 | 26 | 0.064639 |
0 | 27 | 0.064639 |
0 | 28 | 0.03185 |
0 | 29 | 0.066072 |
0 | 30 | 0.034352 |
0 | 31 | 0.03185 |
0 | 32 | 0.066072 |
0 | 33 | 0.064639 |
0 | 34 | 0.064639 |
0 | 35 | 0.064639 |
0 | 36 | 0.064639 |
0 | 37 | 0.064639 |
0 | 38 | 0.064639 |
0 | 39 | 0.064639 |
0 | 40 | 0.064639 |
0 | 41 | 0.064639 |
0 | 42 | 0.064639 |
0 | 43 | 0.066072 |
0 | 44 | 0.03185 |
0 | 45 | 0.064639 |
0 | 46 | 0.064639 |
0 | 47 | 0.064639 |
0 | 48 | 0.064639 |
0 | 49 | 0.064639 |
0 | 50 | 0.066072 |
0 | 51 | 0.066072 |
0 | 52 | 0.064639 |
0 | 53 | 0.064639 |
0 | 54 | 0.064639 |
0 | 55 | 0.066072 |
0 | 56 | 0.064639 |
0 | 57 | 0.066072 |
0 | 58 | 0.066072 |
0 | 59 | 0.064639 |
0 | 60 | 0.064639 |
0 | 61 | 0.064639 |
0 | 62 | 0.064639 |
0 | 63 | 0.005427 |
0 | 64 | 0.064639 |
0 | 65 | 0.064639 |
0 | 66 | 0.066072 |
0 | 67 | 0.066072 |
0 | 68 | 0.066072 |
0 | 69 | 0.064639 |
0 | 70 | 0.064639 |
0 | 71 | 0.064639 |
0 | 72 | 0.064639 |
0 | 73 | 0.034352 |
0 | 74 | 0.064639 |
0 | 75 | 0.064639 |
0 | 76 | 0.064639 |
0 | 77 | 0.034352 |
0 | 78 | 0.066072 |
0 | 79 | 0.064639 |
0 | 80 | 0.064639 |
0 | 81 | 0.03185 |
0 | 82 | 0.064639 |
0 | 83 | 0.064639 |
0 | 84 | 0.064639 |
0 | 85 | 0.066072 |
0 | 86 | 0.064639 |
0 | 87 | 0.066072 |
0 | 88 | 0.066072 |
0 | 89 | 0.064639 |
0 | 90 | 0.064639 |
0 | 91 | 0.064639 |
0 | 92 | 0.064639 |
0 | 93 | 0.064639 |
0 | 94 | 0.061399 |
0 | 95 | 0.064639 |
0 | 96 | 0.066072 |
0 | 97 | 0.064639 |
0 | 98 | 0.064639 |
0 | 99 | 0.066072 |
PRICE rollout pools
Companion data for the code repository zywang0701/PRICE (paper: PRICE — Priced Rollouts and
Inference-time voting-rule Choice under a token budgEt; arXiv link to follow).
Every experiment in the paper is a CPU replay over these logged pools; generation and scoring
are the only GPU stages. python scripts/download_data.py in the code repository places the
files where the experiment chain expects them and verifies them against manifest.json.
| path | content |
|---|---|
cells/<cell>/pools/pool.parquet |
one row per rollout: query id, rollout id, token length, raw and canonical answer, correctness, full text, and the logged scores (phi_conf = P(True), phi_deepconf2, phi_lik, phi_selfcert, …) |
cells/<cell>/pools/queries.parquet |
query text, gold answer, MATH level and subject |
cells/<cell>/pools/embeddings.npz |
last-layer query embedding of the generator |
cells/<cell>/pools/conf_trace2.parquet |
per-rollout confidence trace summaries used by the DeepConf-family baselines |
cells/<cell>/lp*/clusters.parquet, prep*__phi_deepconf2.npz |
label-free answer clusters and the per-query reward-curve prep tables the replay is built from |
overrides/llama/*_ptrue.parquet |
Llama P(True) scores recomputed with one pinned FP16 implementation (generation/ptrue_llama/ in the code repository) |
Cells: qwen25-1.5b_math500 (500 queries × 128 rollouts), qwen25-1.5b_mathtrain (7,425 × 64),
llama32-3b_math500 (500 × 128), llama32-3b_mathtrain (7,425 × 64). Sampling temperature 0.8,
top-p 1.0, at most 2048 new tokens. qwen25-1.5b_mathtrain has no queries.parquet; the chain
reads the shared MATH-train metadata from the Llama cell.
Generators: Qwen/Qwen2.5-1.5B-Instruct (Apache-2.0) and unsloth/Llama-3.2-3B-Instruct
(Llama 3.2 Community License; built with Llama). Questions from MATH (Hendrycks et al., 2021,
MIT) and MATH-500 (Lightman et al., 2023). The pool files also carry score columns the paper does
not use (phi_prm_*, DeepConf v1); they are kept so the files stay byte-identical to the hashes
the experiment manifests record.
- Downloads last month
- 35