Qwen3-8B-Base-Math-TauSFT

Full BF16 Hugging Face weights exported from the final TauSFT checkpoint at step 405.

Training lineage: Qwen3-8B-Base → Qwen3-8B-Base-Math → TauSFT.

Training

• Initialization: Math-stage checkpoint, iteration 300.
• Supervised data: 25,956 tokenized records from Qwen3-32B Tau-bench rollout data (sft_data.jsonl).
• Training: 405 updates, one configured epoch, global batch size 64.
• Optimizer: Adam; learning rate 1e-5 with cosine decay to 1e-6 and 10% warmup.
• Final reported SFT training loss: 0.3738237.

This repository contains the TauSFT stage before subsequent Tau-bench RL. Training data and optimizer states are not included.

Usage

Use a recent version of transformers with Qwen3 support, plus torch, accelerate, and safetensors.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "willhx/Qwen3-8B-Base-Math-TauSFT"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)
messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    do_sample=False,
    eos_token_id=[tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|im_end|>")],
    pad_token_id=tokenizer.pad_token_id,
)
print(tokenizer.decode(outputs[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

The tokenizer and architecture configuration are preserved from Qwen3-8B-Base. For complete Tau-bench interactions, supply the corresponding retail policy, tool schemas, and environment. The example explicitly stops at the assistant turn-ending token as well as the base EOS token.