rmems/agent-trajectory-preference-pairs

Agent Trajectory Preference Pairs

Purpose

Same-context preference pairs for agents: one task, a verified-good trajectory (chosen) and a realistically bad one (rejected), e.g. hallucinated file, skipped test re-run, or premature "done". For DPO.

Record schema

JSONL, one record per line. Fields:

record_id, family, prompt{task, context}, chosen{trajectory, verdict}, rejected{trajectory, verdict, failure_mode}, margin_evidence{chosen_rerun, rejected_rerun}, provenance{produced_at, actors, lineage}

Illustrative example (EXAMPLE ONLY, not real data):

{
  "record_id": "pref-r01-000001",
  "family": "agent-trajectory-preference-pair",
  "prompt": {
    "task": "Fix failing test in utils.py"
  },
  "chosen": {
    "trajectory": [
      "run_tests",
      "read_file utils.py",
      "apply_patch",
      "run_tests -> pass",
      "finish"
    ],
    "verdict": "REPAIRED_VERIFIED"
  },
  "rejected": {
    "trajectory": [
      "read_file helpers.py (does not exist)",
      "apply_patch",
      "finish: tests pass"
    ],
    "verdict": "CLAIM_UNVERIFIED",
    "failure_mode": "skipped_rerun"
  },
  "margin_evidence": {
    "chosen_rerun": "pass",
    "rejected_rerun": "fail"
  }
}

Generation

• Open-weight local models only, run via Ollama: gemma4:12b (writer/solver), qwen3-embedding:8b (embedding dedup), FrogNano-4B (Apache-2.0). No hosted frontier models.
• Produced via rmems/synthetic-factory, with provenance gates (every record names its actors, model digests, and seeds) and execution-evidence gates (claims are re-derived from measured execution, not trusted from model transcripts).

Intended use

Post-training of open-weight coding/agent models: supervised fine-tuning (SFT) on trajectories and preference tuning (DPO). Not intended for pre-training or for evaluating hosted models.

Status

46 records (train: 46), generated deterministically by the rmems/synthetic-factory CLI at commit 5e3b048. Every record passed the factory's replay gate; each carries its lane, generate/replay commands, seed and source record id under provenance.lineage. No LLM was used. Field names follow the schema below, with extra fields where the factory emits more detail.

Limitations

• Current records come from deterministic scripted/programmatic generators (no LLM); trajectories may be narrower and simpler than real-world agent behavior.
• Task sources (e.g. public repos) are likely present in base-model pretraining; contamination checks and program-level eval splits are planned but not yet validated.
• English only. Schema may still change in later releases.

Known limitations (this release)

• Rejected trajectories are scripted failure variants (give_up_on_fault 25, skip_verification 19, skip_confirmation 2), not organic model failures.
• The same gold chosen trajectory is reused across several rejected ones for a task.
• Tool-world lanes only (mcp 19, browser 11, delegation 10, workspace 6); no code-repair pairs because the repair lane has no failing attempts to pair.
• Tool-call args/arguments are stored as JSON-encoded strings (some episodes deliberately pass wrongly typed arguments, which would otherwise break a fixed column type); json.loads them to recover the original object.