Dataset Viewer
Auto-converted to Parquet Duplicate
messages
listlengths
7
78
reward
float64
0
1
session_id
stringlengths
36
36
model
stringclasses
1 value
n_turns
int64
3
39
n_tool_calls
int64
3
38
n_oracle_runs
int64
0
9
last_oracle_score
float64
0
1
ever_scored_1
bool
2 classes
source
stringclasses
1 value
[ { "content": "You are agent9 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order):\n1. program.md — task description and construction\n2. mathlib_hints.md — exact Mathlib lemma names and tactic patterns\n\nAfter those two reads, your IMMEDIATE next action must be Edit on workspace/agent9/Erd...
1
08dc7947-3463-4c75-b486-cc5cd7b50181
claude-opus-4-8
16
15
3
1
true
rrma-erdos-741ii-g1
[ { "content": "You are agent12 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order):\n1. program.md — task description and construction\n2. mathlib_hints.md — exact Mathlib lemma names and tactic patterns\n\nAfter those two reads, your IMMEDIATE next action must be Edit on workspace/agent12/E...
1
09b99749-c228-4f9c-bb87-b5977bb7954f
claude-opus-4-8
9
8
1
1
true
rrma-erdos-741ii-g1
[ { "content": "You are agent4 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order):\n1. program.md — task description and construction\n2. mathlib_hints.md — exact Mathlib lemma names and tactic patterns\n\nAfter those two reads, your IMMEDIATE next action must be Edit on workspace/agent4/Erd...
1
0d8e3fe6-77e0-49e5-83a2-b905c247d673
claude-opus-4-8
16
15
3
1
true
rrma-erdos-741ii-g1
[ { "content": "You are agent1 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order):\n1. program.md — task description and construction\n2. mathlib_hints.md — exact Mathlib lemma names and tactic patterns\n\nAfter those two reads, your IMMEDIATE next action must be Edit on workspace/agent1/Erd...
1
11e08619-5a0f-41d0-bd75-9a88eb581fa1
claude-opus-4-8
10
9
1
1
true
rrma-erdos-741ii-g1
[ { "content": "You are agent11 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order):\n1. program.md — task description and construction\n2. mathlib_hints.md — exact Mathlib lemma names and tactic patterns\n\nAfter those two reads, your IMMEDIATE next action must be Edit on workspace/agent11/E...
1
345857e5-9de1-4432-929f-3fee59ae4cde
claude-opus-4-8
19
18
5
1
true
rrma-erdos-741ii-g1
[{"content":"You are agent9 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order)(...TRUNCATED)
1
3f046149-4b17-40bf-a6f7-c4091c2fff3f
claude-opus-4-8
15
14
2
1
true
rrma-erdos-741ii-g1
[{"content":"You are agent0 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order)(...TRUNCATED)
0
3f3bbdea-4893-476b-bd1d-f12262e2cf1c
claude-opus-4-8
3
3
0
null
false
rrma-erdos-741ii-g1
[{"content":"You are agent3 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order)(...TRUNCATED)
1
4424f4c8-6710-49d3-8ac8-cc7a83a7f143
claude-opus-4-8
25
24
3
1
true
rrma-erdos-741ii-g1
[{"content":"You are agent1 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order)(...TRUNCATED)
1
4af03bb0-b8fb-469d-b02a-3f73d1b6e10b
claude-opus-4-8
25
24
9
1
true
rrma-erdos-741ii-g1
[{"content":"You are agent3 for a Lean 4 proof domain.\n\nREQUIRED first reads (ONCE, in this order)(...TRUNCATED)
1
4c8e0d30-5aed-4f0e-bcd8-4ca749524b37
claude-opus-4-8
19
18
5
1
true
rrma-erdos-741ii-g1
End of preview. Expand in Data Studio

Erdős #741(ii) — Lean 4 Opus Agent Traces

39 Claude Opus agent sessions attempting Erdős problem #741(ii) in Lean 4 (G1 rung: build the proof from an NL construction description).

v2 (2026-06-10) — label correction. The original upload labeled all 39 traces reward=1.0; the labeler matched the text SCORE=1.0 anywhere in the conversation, including the worker prompt. Rewards are now anchored to genuine oracle output (line-anchored SCORE= adjacent to SORRY_COUNT:/BUILD_EXIT:/STATUS: inside a tool result), with reward = 1.0 iff the session's last oracle score is 1.0. Corrected: 31 of 39 proved; of the other 8, 6 never ran the oracle (session cap / API errors) and 2 ran it but never reached 1.0. Chat structure normalized (strict role alternation, empty <thinking> removed).

Task

Prove that there exists a set of integers that is both an additive basis (every n≥4 is a sum of two elements) and unsplittable (every partition has a non-syndetic half). Construction: tower on powers of 5.

Traces

Each row is one complete agent session:

  • messages: list of {role, content} turns (user prompt + tool calls + oracle feedback + fixes)
  • reward: 1.0 = session's final genuine oracle run scored 1.0 (Lean compiled, 0 sorry), else 0.0 — 31/39
  • n_oracle_runs, last_oracle_score, ever_scored_1: oracle audit fields
  • model: claude-opus-4-8
  • n_turns, n_tool_calls: session stats
  • Tool calls inlined as <tool_call> / <tool_result> markers
  • Thinking blocks preserved as <thinking>

Signal

Hard verifiable reward — Lean compiler accepts or rejects, no fuzzy grading. Traces show the edit→compile→diagnose→fix loop that amortizes formal-language friction.

Usage

from datasets import load_dataset
ds = load_dataset("vincentoh/erdos741ii-lean4-opus-traces", split="train")
proved = ds.filter(lambda x: x["reward"] == 1.0)  # 31 traces
Downloads last month
93