Datasets:
messages list | enable_thinking bool | id string | source string | sample_index int64 | step_idx int64 | num_steps int64 | label string | correct bool | loss_weight float64 | reward float64 |
|---|---|---|---|---|---|---|---|---|---|---|
[
{
"role": "user",
"content": "# Role\nYou are an expert evaluator acting as a process reward model (PRM). You are given a\nproblem and a step-by-step solution that may be only a **partial trace** -- a prefix\nof a longer reasoning chain that can stop in the middle of a step. Judge whether the\nreasoning sho... | false | bb520714cd137b5b | polaris | 0 | 4 | 37 | yes | true | 1 | 0.875 |
[
{
"role": "user",
"content": "# Role\nYou are an expert evaluator acting as a process reward model (PRM). You are given a\nproblem and a step-by-step solution that may be only a **partial trace** -- a prefix\nof a longer reasoning chain that can stop in the middle of a step. Judge whether the\nreasoning sho... | false | bb520714cd137b5b | polaris | 0 | 12 | 37 | yes | true | 1 | 0.875 |
[{"role":"user","content":"# Role\nYou are an expert evaluator acting as a process reward model (PRM(...TRUNCATED) | false | bb520714cd137b5b | polaris | 0 | 16 | 37 | yes | true | 1 | 0.875 |
[{"role":"user","content":"# Role\nYou are an expert evaluator acting as a process reward model (PRM(...TRUNCATED) | false | bb520714cd137b5b | polaris | 0 | 20 | 37 | yes | true | 1 | 0.875 |
[{"role":"user","content":"# Role\nYou are an expert evaluator acting as a process reward model (PRM(...TRUNCATED) | false | bb520714cd137b5b | polaris | 0 | 28 | 37 | yes | true | 1 | 0.875 |
[{"role":"user","content":"# Role\nYou are an expert evaluator acting as a process reward model (PRM(...TRUNCATED) | false | bb520714cd137b5b | polaris | 1 | 12 | 55 | yes | true | 1 | 0.875 |
[{"role":"user","content":"# Role\nYou are an expert evaluator acting as a process reward model (PRM(...TRUNCATED) | false | bb520714cd137b5b | polaris | 1 | 16 | 55 | yes | true | 1 | 0.8125 |
[{"role":"user","content":"# Role\nYou are an expert evaluator acting as a process reward model (PRM(...TRUNCATED) | false | d9a6c07fb72aec89 | polaris | 0 | 12 | 38 | yes | true | 1 | 0.5 |
[{"role":"user","content":"# Role\nYou are an expert evaluator acting as a process reward model (PRM(...TRUNCATED) | false | d9a6c07fb72aec89 | polaris | 0 | 24 | 38 | yes | true | 1 | 0.625 |
[{"role":"user","content":"# Role\nYou are an expert evaluator acting as a process reward model (PRM(...TRUNCATED) | false | 3cd039cbbf4fda4c | polaris | 0 | 0 | 45 | yes | false | 1 | 0.3125 |
MC-value PRM dataset — context mode: full_cot
Data for training an in-context / policy-conditioned Monte-Carlo value PRM. Each row's query is a
partial reasoning prefix; the target reward = V = P(correct | prefix), the Monte-Carlo value estimated
from branched Qwen3.5-4B rollouts on Polaris math problems. The user prompt additionally carries a
"# Other attempts by the same model at this problem" block — the ablation variable.
Context for this variant: Up to K=4 OTHER attempts' full chain-of-thought (reasoning), middle-truncated (head+middle+tail with '[... omitted ...]' markers). No correctness labels.
Filtered to uncertain prefixes: only rows with MC value V ∈ [0.1, 0.9] are kept — the signal region where the value function must actually discriminate. (V is otherwise strongly bimodal at 0/1: ~53% of prefixes are V=1 and ~17% are V=0, which carry little training signal.) MC values come from the offline math_verify + judge relabeling of the generate-only rollouts.
All variants (none, full_cot, solution, solution_label, memento) are built from the SAME prefixes /
MC values / problem-id split, so the only difference is the context block (see the sibling repos
asingh15/prm-mc-value-context-*). Context excludes the prefix's own base sample (no leakage) and is budgeted so
the query prefix + verdict always fit the context window.
Columns
messages: chat turns —user= the judge prompt (role + rules + problem + context + partial solution),assistant= a single"yes"placeholder verdict token (the trained readout position).reward: float in [0.1, 0.9] — the soft MC value target V (P(yes) regresses to this).correct: bool = V >= 0.5.id: problem id ·sample_index: which base rollout the prefix was cut from ·step_idx/num_steps: the prefix cut ·source:polaris·label:"yes"·loss_weight: 1.0 ·enable_thinking: false.
Splits
train = 21996 rows · validation = 876 rows (held out by problem id).
Built by inference/curate/build_prm_mc_context.py --context-mode full_cot.
- Downloads last month
- 31