Model Card Template — tau2-bench SFT runs

• base_model — HF id, e.g. unsloth/Qwen3.5-35B-A3B
• model_family — qwen3.5 | gemma4 | …
• domain — e.g. banking_knowledge
• version — e.g. v6
• training — dict with framework, framework_version, hardware, runtime, epochs, steps, seq_len, lr, lora_rank, lora_alpha, optim, lr_scheduler, warmup, data_file
• repo_files — list of (name, size_human, note) tuples
• inference — dict with vllm_command, tool_parser, reasoning_parser
• known_issues — list of (title, body) tuples (rendered as nested sections; can be empty)
• eval — list of (model, pass_rate, note) rows
• eval_command — the literal tau2 run command
• git — dict with commit, branch, repo_url (e.g. https://github.com/Monte-Inc/tau2-banking-sft)

qwen3.6-35b-a3b-banking-sft-v9

Training

• Framework: Unsloth + TRL
• Config: training/run_qwen_sft.py in tau2-banking-sft@adc8c0e35076d078494008976d6bd0c40f58fa12
• Hardware: NVIDIA GH200 480GB (101 GB)
• Runtime: 21.2 min (227 steps, 1 epoch)
• LoRA: rank 32, alpha 64, dropout 0
• Targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, gate_up_proj
• Hyperparams: lr 0.0002, cosine schedule, warmup 22, optim adamw_8bit, seq_len 8192
• Final loss: 0.5025

Repo contents

• adapter_model.safetensors (see HF) — PEFT LoRA weights
• adapter_config.json (small) — r=32, target_modules + target_parameters
• chat_template.jinja (small) — qwen3.6 native template
• tokenizer.json + tokenizer_config.json (small) — qwen3.6 tokenizer
• trainer_state.json (small) — step-by-step loss + grad-norm + LR history
• training_args.json (small) — exact SFTConfig used

Inference setup

Merge the adapter

Serve with vLLM

Baselines on banking (gpt-5.2 user sim, terminal_use)

tau2-bench eval command

Source code