pandeyankit84/autoscientist-toolcaller-dataset

AutoScientist Tool-Calling Dataset

A curated function-calling / tool-use dataset for the Adaption AutoScientist Challenge. Its distinguishing feature is a large slice of hard negatives and reliability-focused cases — where the correct behavior is not a plain tool call.

Adaptive Data quality (real): on the fixed set (c4923b7f…, graded on 1,000 of 2,440 rows under the free-tier cap) the platform reported 7.0 → 8.1, +15.7%, grade C → B — now confirmed by a completed, uncapped run on the current-generation set (bea4a581…, 5,133/5,157 rows, 2026-07-05): C → B, +15.7%, completion quality +31.5%, percentile 8.4 → 31.5; an earlier 250-row run (a99c0c96…, completed) scored 8.0 → 8.8, +10.0%, grade B. This is the platform's dataset-quality grade — the data-centric "measurable improvement." Full dataset lineage (incl. the cleaned d92279d3… set staged for submission) is in docs/AUTOSCIENTIST_USAGE.md.

Composition

Each row is one example (canonical format in autoscientist_toolcaller/format_utils.py):

{"tools": [ {"name","description","parameters"} ],
 "query": "...",
 "answer": {"type": "tool_call"|"refuse"|"clarify", "calls": [...], "content": "..."},
 "meta": {"source": "toolace|hard_negative|multiturn|schema_drift|multilingual|format_twin|masked_twin|agentic|env", "hn_kind": "...", "mt_kind": "...", "sd_kind": "...", "lang": "en|es|fr|hi|hi-rom", "doc_format": "json|python|xml|compact", "pair_id": "..."}}

Splits: train / val / test, plus test_novel (examples using tools never seen in training — a generalization test). 7,566 examples in the published set (HF + Kaggle) — 7,323 across train/val/test + 243 in the test_novel holdout, spanning 7,315 unique tools for broad generalization; platform-graded snapshots were smaller (see the lineage table in docs/AUTOSCIENTIST_USAGE.md). Every tool_call gold is schema-validated at build by a correct-by-construction drop-guard (0 schema-invalid golds shipped; see stats.json:schema_invalid_dropped). Per-source/kind counts and a mix block (intended-vs-realized shares + a mix_ok guard) are in stats.json.

Slices (counts across all splits):

• positives — real tool-call examples curated from ToolACE (schema-validated), plus a small slice of execution-verified examples from deterministic tool environments (autoscientist_toolcaller/envs.py).
• hard negatives — no_tool → refuse (644, held at ~8.8% of the total set, the research optimum); missing_arg / ambiguous → clarify; over_refusal → must call (hedged-but-satisfiable requests, the counterweight to refusal bias); partial_parallel → two calls (call completeness).
• multi-turn — miss_param, miss_func, long_context (BFCL v3/v4 style), plus verified multi-call env trajectories (2–3 order-independent calls).
• schema-drift — a tool's schema changed under the model: add_required / retype_enum (→ clarify) and rename (→ remap to the new field). Every drift gold is validated against its drifted schema.
• format-invariance twins — the same example with its tool docs re-rendered as Python signatures / XML / a compact list (meta.doc_format), gold identical — targets BFCL-v4's format sensitivity (tool-tuned models notoriously overfit one documentation format). Twins share a pair_id with their source, so a twin can never land in a different split than its source.
• masked twins — Hammer-style function masking: neutral func_i/arg_j names with descriptions kept and the gold call renamed consistently (re-validated) — kills naming-convention shortcuts.
• fault-recovery trajectories — agentic rollouts where a scripted transient tool error (503 / 429 / timeout / malformed payload) interrupts a step and the gold continuation is to retry the same call (failure realism; the env state provably doesn't advance on the fault).

Exact per-kind counts and realized-vs-intended shares are in stats.json (mix block). A companion preference set (pref.jsonl) includes execution-labeled DPO pairs (chosen = verified call, rejected = checker-proven-wrong).

Decontamination: every training query is checked (n-gram + embedding) against public BFCL/ToolACE-style probes and dropped on overlap; a contamination block is recorded in stats.json.

Platform-enhanced variant. Adaptive Data (data-centric) rewrites each row into a richer enhanced_prompt / enhanced_completion — in the completed run it revised 100% of completions (e.g. flagging both an invalid enum and a missing required argument where the original flagged only one). Fetch it with python -m autoscientist_toolcaller.fetch_improved --dataset-id <id> to get a clean enhanced prompt/completion file to train on; this is the artifact behind the platform's quality-grade gain.

Data-quality audit. An adversarial audit of the build pipeline caught the refuse/clarify moat being generated and then silently discarded by query-only dedup (no_tool had collapsed to 8 rows, miss_param to 1). The generators, the slice-mixing math, and dedup were fixed, and regression-tested; the counts above are the post-fix set. Full write-up: docs/DATA_QUALITY_AUDIT.md.

Provenance & licensing

Positives are derived and curated from Team-ACE/ToolACE (Apache-2.0). All hard-negative, multi-turn, and schema-drift examples are original, synthesized from the real tool schemas in the corpus (template-based, seeded, reproducible). Every example is schema-validated; the set is deduplicated (MinHash + semantic) with train/val/test leakage removed. Released under Apache-2.0.

Intended use

Supervised fine-tuning (and preference tuning) of function-calling models that must reliably decide whether and how to call tools — including safe abstention.

Limitations

• Primarily English; a matched-twin multilingual slice (~175 rows: es/fr/hi/hi-rom, shared pair_id per twin set) enables a cross-language Δaccuracy — coverage beyond these languages is future work.
• Synthetic hard-negative/multi-turn phrasing is template-derived (labels correct by construction; wording is less varied than fully natural queries).

Reproduction

python -m autoscientist_toolcaller.build_dataset --config config.yaml (repo: https://github.com/ankit25bcs10610/adaption-ai-lab)