shantomathew's picture
ToolTune datasets, trajectories and evaluation logs (Lexsi Track 3)
d2a4b12 verified
|
Raw History Blame Contribute Delete
1.96 kB
metadata
license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - agenttune
  - auditkit
  - tool-calling
  - lexsi-hackathon
pretty_name: ToolTune Tool-Calling Tasks, Trajectories and Evaluation Logs

ToolTune — datasets & logs (Lexsi Labs AI Safety & Alignment Hackathon, Track 3: Agentic Fine-tuning)

All datasets, trajectories and evaluation logs for the ToolTune submission by Shanto Mathew (AgentMart), chaining AgentTune (post-training trajectories) + AuditKIT (evaluation) for reliable tool calling, targeting Cohere Labs open-weights models.

Contents

  • tasks.json — 170 deterministic tool-calling tasks (130 train / 40 held-out test; prompts and entities fully disjoint across the split) across 5 tools: search, calculator, calendar_lookup, price_lookup, weather_lookup.
  • sft_agenttune.jsonl — 130 SFT rows in AgentTune's native format (as emitted by EventLog.as_dataset_rows('sft'): messages / segment_weights / loss_mask).
  • trajectories.jsonl — 130 trajectory records in AgentTune TrajectoryDataset format; verified loadable by AuditKIT (auditkit.load_agenttune → 130 samples, 130 episodes).
  • metrics.json — the real before/after evaluation from the executed pipeline run (AuditKIT 1.1.11).

Headline results (held-out test, n=40)

Metric Baseline ToolTune policy
Tool-selection accuracy 0.75 1.00
Argument exact-match 0.15 1.00
Task success rate 0.15 1.00
Avg calls / task 1.275 1.00
AuditKIT tool_call_f1 (exact) 0.15 1.00
AuditKIT redundant_tool_calls 0.1375 0.00

Scope note: synthetic single-step tasks; the split is entity/value-disjoint but template wording families are shared, so this is in-distribution generalisation to unseen entities — not a BFCL-style claim. Full write-up: see the model repo shantomathew/agentmart-tooltune-policy and the code repo linked in the hackathon submission.