Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

GLM-5.3-Flash Function-Calling Dataset (work in progress)

Synthetic function-calling traces generated with zai-org/GLM-5.3-Flash as the teacher, called through HF Inference Providers (novita → together → fireworks-ai → deepinfra fallback chain).

Status: awaiting the user's 530-agent list and a compute budget grant. Nothing has been generated yet.

Tool schema source

The tool universe is the user's 530-agent list, mounted into the generation job at /data/agents.json. Expected format:

[
  {
    "name": "search_flights",
    "description": "Search available flights between airports",
    "parameters": {
      "type": "object",
      "properties": {"origin": {"type": "string"}, "destination": {"type": "string"}},
      "required": ["origin", "destination"]
    }
  }
]

The generator validates this up-front and fails fast on missing keys or duplicate names. 4–12 agents are sampled per conversation so the teacher must discriminate among similar tools.

Output format (per row, JSONL)

  • id — sample id
  • scenario_type — single | multi | refusal
  • scenario_kind — the request archetype the teacher was asked to write
  • expected_tools — tool names the teacher said a good assistant should call
  • tools — OpenAI-style function definitions actually offered to the teacher
  • messages — OpenAI chat format: system + user → assistant tool_calls (arguments as JSON strings) → tool results (simulated by the teacher) → follow-up calls or final answer. Hallucinated tool names and unparseable arguments are discarded at generation time.

Generation

  • generate.py — two-stage pipeline: (1) the teacher writes a realistic user request given a sampled toolset, (2) the teacher answers it with the tools offered, optionally with simulated tool results and follow-up calls (multi-turn, depth ≤ 2).
  • reasoning_effort="low" to avoid paying for max-effort thinking tokens; temperature 0.7–0.9 for diversity; seeded RNG for reproducibility.

Reproduce

python generate.py --agents /data/agents.json --output /data/output/pilot.jsonl \
  --n 1000 --mode mix --seed 42 --concurrency 8 \
  --repo-id tjmorrisshow/glm-5.3-func-calling
Downloads last month
44