glm-5.3-flash-funcgen

Full generator script for the synthetic function-calling dataset Offlin33er/glm-5.3-flash-function-calling.

What it produces

Rows with messages (OpenAI tool-use format) + tools (1–4 function schemas) + category, across 5 categories (single_call, parallel_calls, multi_turn, error_handling, no_tool) and 20 domains. Every row passes a validator before being written.

Run it

pip install "huggingface_hub>=1.32,<2" "datasets==5.0.1"
export HF_TOKEN=...            # a token with Inference Providers access + write access

python gen_dataset.py --dry-run                     # self-test, no API calls
python gen_dataset.py --n 5 --out smoke.jsonl       # 5-sample smoke
python gen_dataset.py --n 500 --out dataset.jsonl \
    --push --repo-id <your-namespace>/<dataset-name>  # full run + push

Notes:

• Model: zai-org/GLM-5.3-Flash via HF Inference Providers; auto-routing first, then novita / together / fireworks fallback, with retry backoff.
• On huggingface_hub >= 1.x, provider is a constructor argument to InferenceClient, not a chat_completion() keyword — passing it there is the most common failure mode.
• Production run reference: 500/500 ok, 0 failed, ~7 min on a 2-vCPU machine at 10-way concurrency.
• Sample latency is provider-dependent (~90 s when cold, ~10 s once warm).