Dataset Viewer
Auto-converted to Parquet Duplicate
arm
large_stringclasses
1 value
row
int64
0
372
input
large_stringclasses
373 values
output
large_stringlengths
1.58k
36.4k
gts
large_stringclasses
5 values
score
float64
-0.08
0.99
step
int64
0
40
reward
float64
-0.08
0.99
acc
float64
0
1
num_explores
float64
1
8
cost
float64
0
6.58
has_answer
float64
0
1
penalty
float64
0.01
0.08
utility
float64
-0.08
0.99
discovery
float64
0
1
integration
float64
0
1
grader_gap
float64
0
1
subset
float64
0
1
_atlas_process_enabled
float64
1
1
_atlas_process_mode_code
float64
2
2
_atlas_process_potential_stop
float64
0
0
potential_stop
float64
0
0
potential_defined
float64
0
0
_atlas_process_num_explores
float64
1
8
_atlas_process_did_submit
float64
1
1
_atlas_process_tool_action_count
float64
2
9
_atlas_process_base_total
float64
-0.08
0.99
_atlas_process_terminal_reward
float64
0
1
_atlas_process_explore_reward_1
float64
-0.01
-0.01
_atlas_process_explore_reward_2
float64
-0.01
0
_atlas_process_explore_reward_3
float64
-0.01
0
_atlas_process_explore_reward_4
float64
-0.01
0
_atlas_process_explore_reward_5
float64
-0.01
0
_atlas_process_explore_reward_6
float64
-0.01
0
_atlas_process_explore_reward_7
float64
-0.01
0
_atlas_process_explore_reward_8
float64
-0.01
0
data_source
large_stringclasses
2 values
shaped
0
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidates exist yet, so there is nothing to judge. Call the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: B - Confidence: 0.99 - Approach: Apply the energy-time uncertainty principle to find the natu...
B
0.97
0
0.97
1
3
0.194159
1
0.03
0.97
1
1
0
1
1
2
0
0
0
3
1
4
0.97
1
-0.01
-0.01
-0.01
0
0
0
0
0
atts_gpqa
shaped
1
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No evidence exists yet, so there is nothing to weigh. Dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: B - Confidence: 0.92 - Approach: Tracing each reaction step to count carbons in the final ...
B
0.99
0
0.99
1
1
0.040133
1
0.01
0.99
1
1
0
1
1
2
0
0
0
1
1
2
0.99
1
-0.01
0
0
0
0
0
0
0
atts_gpqa
shaped
2
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidate has been revealed, so there is nothing to judge yet. Spend the first explore call on a solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: +∞ (D) - Confidence: 0.99 - Approach: Compute the area integral...
D
0.99
0
0.99
1
1
0.071989
1
0.01
0.99
1
1
0
1
1
2
0
0
0
1
1
2
0.99
1
-0.01
0
0
0
0
0
0
0
atts_gpqa
shaped
3
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidate has been revealed, so there is nothing to judge. Dispatch a solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: D - Confidence: 0.99 - Approach: Direct computation of expectation values using the given s...
D
0.97
0
0.97
1
3
0.257916
1
0.03
0.97
1
1
0
1
1
2
0
0
0
3
1
4
0.97
1
-0.01
-0.01
-0.01
0
0
0
0
0
atts_gpqa
shaped
4
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidates revealed yet, so there is nothing to compare and no direction I can form. Dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: B - Confidence: 0.99 - Approach: Analyze which Maxwell's...
B
0.97
0
0.97
1
3
0.114917
1
0.03
0.97
1
1
0
1
1
2
0
0
0
3
1
4
0.97
1
-0.01
-0.01
-0.01
0
0
0
0
0
atts_gpqa
shaped
5
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidates exist yet, so there is nothing to compare. Dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: B - Confidence: 0.88 - Approach: Used the Ksp of Fe(OH)₃ (= 10⁻³⁶) to find the minimum ...
B
0.97
0
0.97
1
3
0.464207
1
0.03
0.97
1
1
0
1
1
2
0
0
0
3
1
4
0.97
1
-0.01
-0.01
-0.01
0
0
0
0
0
atts_gpqa
shaped
6
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> Nothing has been revealed yet, so there is no evidence to weigh. Dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: D - Confidence: 0.99 - Approach: Find the eigenvector of the spin-1/2 operator ...
D
0.99
0
0.99
1
1
0.037915
1
0.01
0.99
1
1
0
1
1
2
0
0
0
1
1
2
0.99
1
-0.01
0
0
0
0
0
0
0
atts_gpqa
shaped
7
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidate has been revealed, so there is nothing to judge yet. Dispatch the first solver and read what it produces. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: D - Confidence: 0.99 - Approach: Convert polar poten...
D
0.99
0
0.99
1
1
0.045078
1
0.01
0.99
1
1
0
1
1
2
0
0
0
1
1
2
0.99
1
-0.01
0
0
0
0
0
0
0
atts_gpqa
shaped
8
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidates yet; there is nothing to weigh. Dispatch a solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: A - Confidence: 0.97 - Approach: Apply the threshold condition for pair production using the invariant mas...
A
0.98
0
0.98
1
2
0.069352
1
0.02
0.98
1
1
0
1
1
2
0
0
0
2
1
3
0.98
1
-0.01
-0.01
0
0
0
0
0
0
atts_gpqa
shaped
9
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> Nothing has been revealed yet, so there is no evidence to manage. Spend the first call on a solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: D - Confidence: 0.82 - Approach: Systematic analysis of single and doubl...
D
0.98
0
0.98
1
2
0.146794
1
0.02
0.98
1
1
0
1
1
2
0
0
0
2
1
3
0.98
1
-0.01
-0.01
0
0
0
0
0
0
atts_gpqa
shaped
10
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No evidence yet. Dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: A - Confidence: 0.92 - Approach: Analyzing the symmetry elements of each molecule to identify which has exactly C3h symmetry (C...
B
0.96
0
0.96
1
4
0.424697
1
0.04
0.96
1
1
0
1
1
2
0
0
0
4
1
5
0.96
1
-0.01
-0.01
-0.01
-0.01
0
0
0
0
atts_gpqa
shaped
11
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No evidence yet; I dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: D - Confidence: 0.95 - Approach: Using the mass function from radial velocity amplitudes and Kepler's third law to compute tot...
D
0.97
0
0.97
1
3
0.11571
1
0.03
0.97
1
1
0
1
1
2
0
0
0
3
1
4
0.97
1
-0.01
-0.01
-0.01
0
0
0
0
0
atts_gpqa
shaped
12
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidate has been revealed, so there is nothing to weigh. Dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: A - Confidence: 0.55 - Approach: Analyzing each statement against known organometa...
B
0.98
0
0.98
1
2
0.266919
1
0.02
0.98
1
1
0
1
1
2
0
0
0
2
1
3
0.98
1
-0.01
-0.01
0
0
0
0
0
0
atts_gpqa
shaped
13
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidates yet; dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: A - Confidence: 0.92 - Approach: Analyzing density scaling for rocky planets with Earth-like composition using mass-radius re...
A
0.97
0
0.97
1
3
0.108029
1
0.03
0.97
1
1
0
1
1
2
0
0
0
3
1
4
0.97
1
-0.01
-0.01
-0.01
0
0
0
0
0
atts_gpqa
shaped
14
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidate has been revealed, so there is nothing to compare and no legitimate direction to take. Dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: A - Confidence: 0.85 - Approach: Analyzing e...
A
0.96
0
0.96
1
4
0.314055
1
0.04
0.96
1
1
0
1
1
2
0
0
0
4
1
5
0.96
1
-0.01
-0.01
-0.01
-0.01
0
0
0
0
atts_gpqa
shaped
15
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No candidates exist yet, so there is nothing to weigh. Dispatch a solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: D - Confidence: 0.99 - Approach: Applied 4-momentum conservation for pion decay at rest using rela...
D
0.98
0
0.98
1
2
0.105863
1
0.02
0.98
1
1
0
1
1
2
0
0
0
2
1
3
0.98
1
-0.01
-0.01
0
0
0
0
0
0
atts_gpqa
shaped
16
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> No evidence yet. Dispatch the first solver on this four-step sequence and read its answer. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: D - Confidence: 0.82 - Approach: Tracking stereochemistry and regiochemistry thro...
C
0.95
0
0.95
1
5
0.896565
1
0.05
0.95
1
1
0
1
1
2
0
0
0
5
1
6
0.95
1
-0.01
-0.01
-0.01
-0.01
-0.01
0
0
0
atts_gpqa
shaped
17
system You are a meta-reasoning orchestrator. You manage a growing set of candidate solutions for a problem. Each explore call dispatches a solver and reveals one candidate result (answer, confidence, approach, reasoning). You cannot solve the problem yourself; you decide when to explore further and when to submit. Y...
<think> Nothing has been revealed yet, so there is no evidence to weigh. Dispatch the first solver. </think> <tool_call>{"name": "mcp__delegated__explore", "arguments": {} }</tool_call>user <tool_response> Candidate #1 recorded. - Answer: C - Confidence: 0.88 - Approach: Applying knowledge of ChIP-seq crosslinking che...
C
0.99
0
0.99
1
1
0.061857
1
0.01
0.99
1
1
0
1
1
2
0
0
0
1
1
2
0.99
1
-0.01
0
0
0
0
0
0
0
atts_gpqa
End of preview. Expand in Data Studio

ATLAS report 24: validation trajectories of both arms

Every validation trajectory that the two training runs of report 24 (frozen-prefix potential shaping) dumped, as one ordinary Hugging Face dataset: two splits, one per arm, one row per (validation step, question).

  • shaped: the shaped arm (W&B run 7mabv2p1), validations at steps 0, 5, 10, 15, 20, 25, 30, 35, 40 (3357 rows).
  • baseline: the baseline arm (W&B run 3fez6lk9), validations at steps 0, 5, 10, 15, 20 (1865 rows). The baseline stopped at step 20; its step-20 validation holds the degenerate trajectories described in the experiment's front page (rows 11, 184, 339 are whitespace loops, row 270 a two-call turn).

Each row: arm, row (the line index in the source dump), step, data_source (atts_gpqa or atts_lcb), input (the full prompt), output (the full trajectory text, tool calls and tool responses included), gts (the gold answer), the graded fields score, reward, acc, utility, cost, penalty, num_explores, has_answer, discovery, integration, grader_gap, subset, and the process-reward bookkeeping fields (_atlas_process_*, potential_stop, potential_defined).

Source: the checkpoints/<arm directory>/val_dumps/<step>.jsonl files of the experiment's data root, published in full at https://huggingface.co/datasets/t2ance/atlas-24-frozen-prefix-potential-shaping (front page, tables, logs, exports) and mirrored without checkpoints in the private GitHub repository t2ance/atlas-experiments. This dataset is a reading copy of those dumps and adds nothing to them.

from datasets import load_dataset
ds = load_dataset("t2ance/atlas-24-validation-trajectories")
bad = ds["baseline"].filter(lambda r: r["step"] == 20 and r["row"] in (11, 184, 270, 339))
Downloads last month
21