Qwen3-0.6B-ToolCalling-Claude-4.6-Opus-Distilled-v1

1. SFT on mixed tool-calling + reasoning data
2. DPO to improve call-vs-no-call decisions

Model Summary

• Base model: Qwen/Qwen3-0.6B
• Method: LoRA SFT + LoRA DPO
• Primary goal: improve tool invocation decisions and argument formatting while keeping concise reasoning behavior
• Final artifacts in this repo:
  • lora_adapter/* (adapter weights)
  • merged_safetensors/* (merged full model)
  • merged_gguf/model-f16.gguf

Training Data

• Salesforce/xlam-function-calling-60k
• nvidia/When2Call
• Roman1111111/claude-opus-4.6-10000x
• Crownelius/Opus-4.6-Reasoning-3300x

• SFT train: 91,048
• SFT val: 1,858
• DPO train: 8,550
• DPO val: 450

Training Procedure

Stage 1: SFT

• LoRA: r=64, alpha=128, dropout=0.05, target_modules=all-linear
• Sequence length: 16384
• Epochs: 3
• Per-device batch size: 2
• Gradient accumulation: 16
• LR: 2e-4 (cosine, warmup ratio 0.05)
• bf16 + gradient checkpointing
• Best SFT eval loss: 0.126719 (checkpoint-8500)

Stage 2: DPO

• LoRA: r=32, alpha=64, dropout=0.05, target_modules=all-linear
• beta=0.1, loss_type=sigmoid
• max_length=1024, max_prompt_length=512
• Per-device batch size: 1
• Gradient accumulation: 32
• LR: 5e-7 (cosine, warmup ratio 0.03)
• bf16 + gradient checkpointing
• precompute_ref_log_probs=true, precompute_ref_batch_size=1

Runtime Environment

• GPU: NVIDIA GeForce RTX 5090 (31.37 GiB)
• CUDA runtime (PyTorch): 12.8
• PyTorch: 2.11.0+cu128
• TRL: 1.0.0
• PEFT: 0.18.1
• Transformers: 5.5.1

Evaluation

• Built-in function-calling checks (4 cases):
  • Format correctness: 3/4 (75%)
  • Tool selection: 3/4 (75%)
• When2Call MCQ (sample=100 from test/mcq):
  • Accuracy: 67/100 (67%)

Usage

Transformers (merged model)

LoRA adapter

Limitations

• This model can still over-call tools when user input lacks required slots.
• Built-in check includes only 4 handcrafted cases and should not be treated as a benchmark.
• When2Call score is from a 100-sample quick evaluation, not full test set scoring.
• Outputs may inherit bias and errors from source datasets and synthetic data.

License