arjhinety/OpenGrad-ToolPolicy-Canonical-v2
This is a provenance-preserving canonical candidate corpus. It is a pre-training canonical release, not an empirically selected or recommended training mixture. What this release is OpenGrad ToolPolicy Canonical v2 is a provenance-preserving, model-independent normalization of public tool-use datasets in which every record declares what it supervises. It exists because not every legitimate post-training corpus has the same conversational trajectory shape, and discarding a… See the full description on the dataset page: https://huggingface.co/datasets/arjhinety/OpenGrad-ToolPolicy-Canonical-v2.
This is a provenance-preserving canonical candidate corpus. It is a pre-training canonical release, not an empirically selected or recommended training mixture.
What this release is
OpenGrad ToolPolicy Canonical v2 is a provenance-preserving, model-independent normalization of public tool-use datasets in which every record declares what it supervises. It exists because not every legitimate post-training corpus has the same conversational trajectory shape, and discarding a corpus for lacking an arbitrary turn is a loss of real supervision.
This is the completed v2 release: four sources, 173,237 canonical records, 176 shards. It is distinct from the historical partial snapshot (`OpenGrad-ToolPolicy-Canonical-v2-M0-snapshot`), which is unchanged and still published. The two differ in three measured ways:
- this release includes xLAM/APIGen, which contributed 0 trainable records to the partial snapshot and contributes 56,090 here;
- every record carries a supervision contract, so call-prediction and complete-trajectory targets are validated, rendered and measured separately;
- the partial snapshot's When2Call artifact was an interrupted materialization — it had read 9,162 of the raw file's 15,000 rows and was never finalized — so it published 4,000 records where this build retains 6,505.
Supervision contracts
Two contracts are implemented, and exactly one rule differs between them: whether a terminal tool call requires a future environment response.
A CALL_PREDICTION record is a query + tools -> assistant tool_call example whose supervised target is the call. No tool result is fabricated to make it look like a trajectory, and no placeholder observation is inserted. The contract changes which turns are expected, never which turns exist.
Everything else fails closed under both contracts: undeclared tools, invalid arguments against the effective schema, malformed calls, duplicate call ids, orphaned results, FIFO order violations, and a non-terminal call with no result. A malformed call stays quarantined even when its shape is a valid call-prediction example.
Source manifest
† For xLAM and When2Call this is a Hub commit. For Glaive and ToolACE it is not a Hub revision: it is the SHA-256 of the local input file the adapter read (for ToolACE, the first 16 hex characters of 7a7a6a2c3b1003c789bb…), and it does not match the Hub file's LFS hash either. The Hub revisions are e7f4b6456019f5d8bcb991ef0dd67d8ff23221ac (Glaive) and 6bda777c88d21e5a204703c1ee45597a8fa4f734 (ToolACE): Canonical-v1 records the same input files, by SHA-256, against those revisions, and the Hub history shows neither upstream data file has changed since it was uploaded.
Composition and trainability
Measured at the training boundary by rendering every record, not by counting canonical rows.
The mixture is natural: no supervision kind is reweighted and none is excluded.
What each source actually supervises
- xLAM/APIGen is next-tool-call prediction. It contains no tool-result turn at all, which is why it needs
CALL_PREDICTION; its upstream card documentsanswersas the call to make. - Glaive supplies real multi-turn trajectories: every call is answered by a function-response turn.
- ToolACE retains 20.4% of its canonical records. 8,476 end on an assistant call that no tool message answers. Those are quarantined rather than reclassified, because the available bytes do not establish whether they are intended next-call supervision or truncated trajectories, and guessing would train on incomplete conversations while labelling them deliberate.
- When2Call supervises the decision in prose — decline, ask for the missing detail, or answer directly. It carries no structured call and no tool result, and that is its contribution rather than a defect.
Refusal text under ANSWER labels
The refusal detector used to score model outputs (HEURISTIC_REGEX_v1) finds refusal-shaped supervised targets that carry the decision label ANSWER. Counting single-exchange records only (a multi-turn record's label describes its first exchange, not its last turn), 18,114 of the 173,237 records (10.5%) have one, and all 18,114 are labelled ANSWER:
A record labelled ANSWER whose target declines teaches that answering can look like declining. M0, trained on this corpus, and M1-v2, trained from M0, decline all 1,319 zero-shot GSM8K questions; M1-v1, DPO applied directly to the base model without this corpus, declines 70.7%. The count is descriptive and does not establish that these records caused that behaviour; that would need a retrain without them. When2Call entered M0's training in full, so its count applies exactly; Glaive's and ToolACE's are over release records, some of which M0's renderer quarantined. Source: results/benchmarks/h200/capability_v1/sft_refusal_supervision_audit_canonical_v2.json, which supersedes an audit of the normalization-v1 sources (21,749 of 217,903) that was not this corpus.
What is not included
xLAM's upstream is also access-gated: its records are reconstructed from the published Canonical-v1 derivative, which retains tools[].parameters unmodified. The derivation is representational and verified record by record (59,370/59,370 with zero mismatches); it invents no field and repairs nothing.
Quarantine
The last four rows are the training-boundary drops: 1,231 + 8,476 + 700 + 864 = 11,271 of the 173,237 canonical records, leaving 161,966 trainable. An earlier version of this table gave UNRENDERABLE as 2,095, which counted xLAM's 1,231 a second time.
Quarantine is a result, not a shortfall. A record that cannot be interpreted under its declared contract is reported rather than coerced into training text.
Evaluation boundary
The frozen When2Call held-out namespace (when2call-mcq, when2call-llm-judge, and the preference split) is excluded from this payload and remains separate in the OpenGrad repository. It holds 3,952 records but 3,652 distinct examples, because all 300 LLM-judge rows duplicate MCQ rows. Two held-out examples were adjudicated CONTAMINATED against this corpus and quarantined, leaving 3,650 distinct evaluation examples. Contamination review is complete for this corpus and its evidence artifacts are corpus-scoped.
Provenance and versioning
The release manifest records the OpenGrad commit, source manifest hashes, source revisions, adapter versions, output shard hashes and release filters. The recorded commit contains the code that produced this payload, so the release is reproducible from it: two independent builds at that commit produce an identical manifest and 176 byte-identical shards.
This release is frozen. A future change is a new version with a new identity, never an in-place edit.
Licensing and citations
OpenGrad source code is Apache-2.0. Upstream dataset terms remain source-specific and are documented in source-licenses.md; this release does not relicense upstream data. See CITATIONS.bib for source references.
Responsible use
Use the data in accordance with each upstream source's terms and attribution requirements. Do not infer that a valid tool call demonstrates reliable tool-use policy or task completion. Records supervised under CALL_PREDICTION teach next-call prediction; they are not evidence of tool-result interpretation, multi-turn recovery, or multi-step planning, and should not be described as complete tool-use trajectories.