corelabel/agent-trading

Synthetic Trading Agent Trajectories

Step-annotated, tool-using agent trajectories for a discretionary news-driven trading workflow: scan the feed, form a thesis, size a position, place a stop, revise, exit. Built for SFT, process reward modelling (PRM) and RL against a deterministic simulator with verifiable rewards.

Simulated only. Invented tickers, synthetic prices, synthetic news, a mock broker. No real market data. Rewards score process discipline against a simulator, not profitability. Not investment advice; not validated for use with real capital.

Contents

	v1	v2
Trajectory trees	7,670	8,432
Terminal paths	10,391	12,270
Annotated steps	46,725	58,943
Preference pairs	1,601	2,299
SFT paths	3,988	4,727
Size	5.5 MB	7.1 MB

Structure

trajectories/*.jsonl.zst — one JSON object per line, one tree per line (nodes, sibling_groups, terminal_paths), schema in schemas/trajectory.schema.json. Each agent node carries a step label (optimal…harmful), error tags, a progress delta, a counterfactual action, and an MC value target on branched nodes.

derived/ — five flat views: sft, prm_steps, prefs, rl_tasks (task ids only, no demonstrations), eval_suite.

import json, zstandard as zstd
txt = zstd.ZstdDecompressor().decompress(open(path,'rb').read()).decode()
rows = [json.loads(l) for l in txt.splitlines() if l]

Limitations

No model was called anywhere. Behaviour is scripted (an oracle plus degradation mutators, sampled in v2), thoughts are templated — do not train a reasoning style on this — and annotation is rule-based (rule:annot@1.0), so inter-annotator κ is unmeasured and the rubric reward term is 0. Per-episode P&L is recorded with weight 0 by design. Microstructure is simplified; options are single-leg only.

Gates that pass: schema validity, replay determinism, look-ahead firewall, split disjointness. Gates that fail, with reasons, are in v*/gates_report.json and docs/datasheet-set2-trading.md.

Licence

Apache 2.0