agentmart-tooltune-data / export_policy.py
shantomathew's picture
ToolTune code, article and training notebook (Lexsi Track 3)
1a7314a verified
Raw History Blame Contribute Delete
1.34 kB
#!/usr/bin/env python3
"""Export the fitted LearnedPolicy parameters to JSON for the HF model repo."""
import json, sys
sys.path.insert(0, "/home/hatch/workspace/lexsi-agentmart/src")
sys.path.insert(0, "/home/hatch/workspace/lexsi-agentmart")
from policies import LearnedPolicy
p = LearnedPolicy().fit("/home/hatch/workspace/lexsi-agentmart/data/sft_agenttune.jsonl")
out = {
"artifact": "ToolTune learned agent policy (naive-Bayes tool router + cue/shape argument extractor)",
"fitted_on": "data/sft_agenttune.jsonl (130 AgentTune-format SFT trajectories)",
"n_train": p.n_train,
"tool_counts": dict(p.tool_counts),
"vocab_size": len(p.vocab),
"feat_counts": {t: dict(c) for t, c in p.feat_counts.items()},
"arg_stats": {
f"{tool}.{arg}": {"cues": dict(s["cues"]), "stops": dict(s["stops"]), "shapes": dict(s["shapes"])}
for (tool, arg), s in p.arg_stats.items()
},
"note": "This is the locally evaluated policy artifact from the Lexsi Track 3 pipeline. The Cohere Aya LoRA adapter is produced by train_aya_sft.ipynb (GPU step, not executed in this run) and is not claimed here.",
}
with open("/home/hatch/workspace/lexsi-agentmart/results/policy_weights.json", "w") as f:
json.dump(out, f, indent=1)
print("wrote policy_weights.json, n_train =", p.n_train, "vocab =", len(p.vocab))