edithatogo/qwen3-hermes-strict-toolcall-synthetic-v4

Qwen3 Hermes Strict Tool-Call Synthetic V4

Registry status

• —Registry ID: edithatogo/qwen3-hermes-strict-toolcall-synthetic-v4
• —Family: hermes
• —Repository role: canonical_training_dataset
• —Canonical dataset: `edithatogo/qwen3-hermes-strict-toolcall-synthetic-v4`
• —Operational status: active
• —Rights status: apache-2.0-synthetic
• —Authoritative catalog: `edithatogo/dataset-estate-registry`

Origin and provenance

• —Origin repository: https://github.com/edithatogo/models_lang
• —Upstream source: Repository-authored synthetic examples.
• —This card describes repository identity and relationships; it does not claim completeness beyond published manifests.

Rights boundary

Repository code, dataset compilation, and underlying source-record rights are separate. The registry currently records the rights state as apache-2.0-synthetic. Where rights vary by record or source, downstream users must inspect the published provenance and rights fields.

Existing dataset documentation

This dataset contains the approved cleaned synthetic-only training/evaluation splits for local Hermes-style strict tool-call fine-tuning experiments.

The approved publication source was materialized locally at:

/Volumes/PortableSSD/hermes-evals/datasets/qwen3-v4-synthetic-only-20260526

Dataset Summary

The dataset contains 82 repo-authored synthetic examples covering strict JSON tool-call output, invalid-tool handling, argument correctness, and multi-turn repair formatting.

It is intended as reproducibility companion material for the experimental Qwen3 v4 Hermes strict tool-call adapter:

https://huggingface.co/edithatogo/qwen3-4b-hermes-lora

Splits

Total rows: 82

Unique IDs: 82

Duplicate IDs: 0

Included Sources

Only cleaned synthetic expansion rows are included. Some row metadata preserves +no_think_prompt source suffixes for prompt-variant examples, but those rows come from the same approved synthetic expansion families.

The local seed and mirrored regression rows were excluded from this public dataset scope.

Intended Use

• —Local supervised fine-tuning experiments for strict Hermes-style tool-call output.
• —Regression testing for JSON validity, argument correctness, invalid-tool handling, and multi-turn repair formatting.
• —Reproducibility companion material for the experimental Qwen3 v4 strict tool-call adapter.

Not Intended For

• —Broad claims about BFCL, IFEval, HumanEval, MBPP, production tool-use performance, or general assistant quality.
• —Training without the documented Qwen runtime prompt condition used by the associated local adapter experiments.
• —Dataset contamination studies without reviewing the companion overlap and source audits retained in the project repository.

Provenance And Audit Notes

The rows are repo-authored synthetic examples. The public scope intentionally excludes local seed rows and mirrored regression material.

Project-side publication evidence recorded before upload:

• —source audit: cleaned-synthetic-source-audit.json
• —overlap audit: cleaned-synthetic-overlap-audit.json
• —token audit: cleaned-synthetic-token-audit.json
• —publication dry run: dataset-publication-dry-run-20260526.md
• —scope decision: dataset-publication-scope.md

The cleaned synthetic-only candidate had no duplicate IDs and no held-out user-prompt overlap in the recorded audit.

License

Apache-2.0.

Reconciliation policy

This repository is retained only while its recorded role is distinct from the canonical dataset. Renames, deletion, visibility changes, and replacement of immutable DOI snapshots require explicit human review.