Datasets:

azlanmalikai
/

agent-compass-data

license: apache-2.0
task_categories:
  - text-classification
language:
  - en
tags:
  - agents
  - value-model
  - process-reward-model
  - swe-bench
  - trajectories
size_categories:
  - 100K<n<1M
pretty_name: agent-compass training data

agent-compass-data

Training, development and test records for agent-compass, a calibrated value model for AI agents. Each record is one agent state (task plus trajectory prefix, rendered to text) with labels for up to six typed questions.

Sources: nebius/SWE-agent-trajectories (80,034 runs), nebius/SWE-rebench-openhands-trajectories (67,074 runs) and AgentSuite/tau2-bench-trajectories (8,340 runs). Every run has a measured outcome.

Files

path	questions	records	note
`train.jsonl`, `dev.jsonl`, `test.jsonl`	`p_success`	49,987 / 9,999 / 64,124	v0: the success-only release
`v1/train.jsonl`	all six	27,000	12,000 trajectories per SWE scaffold x 2 prefixes, 3,000 tau2-bench states; 3,360 success/failure pairs linked by `_meta.pair_id`
`v1/dev.jsonl`	all six	11,005	5,000 per SWE scaffold plus 1,000 tau2
`v1/test.jsonl`	all six	64,124	held-out repositories, both SWE scaffolds (same states as v0 test)
`v1/test_tau2.jsonl`	five (no progress)	3,832	held-out tau2-bench tasks, 840 runs, 30 policy models
`stats.json`, `v1/stats.json`			counts, positive rates, token totals

The v1 files carry the refined stuck rule (judge precision about 88%). The released v1 model was trained on an earlier version of that rule.

Record format

{
  "state": "<task> ... </task>\n<policy> ... </policy>\n<hist> ... </hist>\n<recent> ... </recent>",
  "questions": {
    "p_success": {"type": "noul", "instructions": "...", "label": false},
    "stuck":     {"type": "noul", "instructions": "...", "label": false},
    "progress":  {"type": "score", "instructions": "...", "criteria": ["regressed", "no change", "small progress", "big progress"], "label": 1},
    "steps_left":{"type": "score", "instructions": "...", "criteria": ["1 to 5 steps", "6 to 15 steps", "16 to 40 steps", "more than 40 steps"], "label": 2},
    "best_next": {"type": "choice", "instructions": "...", "criteria": {"option_1": "...", "option_2": "..."}, "label": "option_2"},
    "escalate":  {"type": "noul", "instructions": "...", "label": false}
  },
  "_meta": {"id", "traj_id", "task_id", "group_id", "source", "split", "domain", "scaffold", "policy_model",
            "policy_shown", "repo", "prefix_len", "n_steps", "prefix_frac", "outcome", "baseline", "advantage",
            "stuck_reasons", "progress_reasons", "best_next" (candidate-set tier and provenance), "state_tokens",
            "variant", "pair_id" (train only)}
}

Questions are omitted when their label is undefined (for example steps_left on failed runs, best_next when no candidate set exists, progress on tool-use states). The questions map is the System One request shape (noul / choice / score), so a record is also a valid scoring request.

How states are built

Task text, optional policy-model tag (dropped in 30% of training records), older steps compressed to one line each, the last 8 steps in full with truncated observations. Budget 4,096 tokens (Qwen3.5 tokenizer).
Agent thoughts removed. Padded test output collapsed, harness status lines and outcome-revealing submit messages stripped.
Prefixes sampled at several points of each run; records whose last visible action is the final submit are excluded.

Labels

question	source
`p_success`	measured outcome of the run (hidden tests for SWE, task checks for tau2). `_meta.baseline` is the leave-one-out pass rate of the task, `advantage` the outcome minus baseline
`stuck`	rule: repeated near-identical commands, repeated error signature, edit-revert cycles; navigation commands excluded; validated against a judge model
`progress`	rule over the last step: test counts, error counts, new files, patch size, mapped to 4 levels
`best_next`	candidate set from other runs of the same task with a similar prefix (tier `branching`) or hard negatives from failed runs at a similar step (tier `hard_negative`); positive = action whose continuation succeeded more often
`steps_left`	remaining steps of successful runs, binned
`escalate`	failed run on a task whose leave-one-out pass rate is at most half the policy's overall pass rate

Rule definitions and the judge validation are in docs/label_qa_findings.md of the repository.

Splits

By repository hash (SWE: owner/repo; tau2: domain and task slug), 80/10/10, with SWE-bench Verified repositories forced into the held-out test split. No task or repository appears in more than one split.

License

Apache-2.0 for the processing and labels. The underlying trajectories keep the licenses of their sources: the two Nebius datasets are CC-BY-4.0; AgentSuite/tau2-bench-trajectories states no license on its card (tau2-bench itself is MIT).