qwen3.8-9b-hermes-function-calling-v1

Qwen3.8-9B Hermes Function Calling v1

A QLoRA fine-tune of Empero/Qwen3.8-9B (Qwen3.5 hybrid-architecture distill, 9B dense, 262,144-token context) that specializes the model in Hermes-style tool-call envelopes: given a <tools> XML schema in the system prompt, it emits <tool_call><function=NAME><parameter=KEY>VALUE</parameter></function></tool_call>.

• Base model: Empero/Qwen3.8-9B (Apache-2.0)
• Training data: 1,090 tool-call conversations rendered in ShareGPT format, all with <tools> XML schemas in the system prompt (synthetic/curated tool-use examples across ~2,000 tool names and dozens of domains)
• Method: QLoRA 4-bit, rank 16 / alpha 16, dropout 0, targets q/k/v/o/gate/up/down_proj, lr 2e-4, max seq 4096, batch 2 x grad-accum 4 (effective 8), 2 epochs = 274 steps, cosine schedule
• Trainable params: 29,097,984 / 8,980,926,976 (0.32%)
• Embeddings: frozen (tool tokens are EOS-anchored copies, not trained)
• Vocabulary: resized 248044 → 248079 (2 added tool tokens + base additions)
• Weights: BF16 full merge (12 shards), plus a ROCmFPX quantized GGUF (see the GGUF companion repo)

What this model does

The fine-tune teaches the format: when the system prompt carries function signatures inside <tools> </tools> XML tags (the training preamble), the model reliably produces the exact Hermes envelope with correct function names and parameters, with no prose around it.

Format-fidelity probe (45 fresh training-distribution rows, temp 0, identical prompts for both models):

The base model answers the same prompts in native Qwen JSON; the fine-tune emits the trained Hermes envelope.

Honest limitations

• No trigger generalization. Evaluated on 45 held-out real agent trajectories (ambiguous queries, real tool names like terminal, search_files, patch — none of which appear in training), both the fine-tune and the base score 0/45. A 0.32% LoRA on 1,090 examples does not teach when to decide to call a tool, especially for unseen tool names and ambiguous user intent.
• Parity, not gains, in OpenAI-style tool-calling paths. Through an OpenAI-compatible tools parameter (the path most serving stacks use), fine-tune and base perform identically — both fire valid calls at ~55–58 tok/s on this rig; the fine-tune shows no regression and no measurable gain there, because the chat template already formats tool calls server-side and the base was already competent.
• The trained envelope only pays off in raw-text tool-call paths: prompts that embed the schema as <tools> XML and parse <tool_call> envelopes from the output (e.g. Hermes's mined-trajectory format, custom parsers).

Reproducibility

All scripts that produced this model are in scripts/:

The scripts carry the author's machine paths (F:/models/..., C:/AI/...) and run on this rig's ROCm stack (Unsloth 2026.8.18, torch 2.11 ROCm, RX 7700 XT); adjust paths for your environment.

Convert/quantize: llama-rocmfpx fork, convert_hf_to_gguf.py --outtype bf16 → llama-quantize Q4_0_ROCMFP4_FAST (see GGUF repo).

Quantized version

Q4_0_ROCMFP4_FAST GGUF (ROCmFPX format, for AMD RDNA3 kernels) is published in the companion repo. Note: ROCmFPX quants target AMD ROCm inference (ROCMFP4 kernels); for portable use, convert from the BF16 merge here.

Acknowledgements

Base model and license inherited from Empero/Qwen3.8-9B (Apache-2.0). Training data derived from Hermes agent trajectories; tool schemas rendered in the Hermes <tools> XML format.