Qwen3-8B-Base-TauSFT-Tau

RL checkpoint of Qwen3-8B-Base for tau-bench agentic tool-use, trained with slime (Megatron-LM training + SGLang rollout).

Pipeline: Qwen3-8B-Base → SFT on tau-bench trajectories (TauSFT) → RL on tau-bench (Tau).

This is the checkpoint at iteration 180, converted from Megatron torch_dist format to Hugging Face safetensors (bf16).

Training setup

Base model	Qwen/Qwen3-8B-Base
Framework	slime (Megatron-LM + SGLang)
Task	tau-bench (tool-use dialogues)
User simulator	zai-org/GLM-4.7-Flash via an OpenAI-compatible endpoint
Precision	bfloat16

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "willhx/Qwen3-8B-Base-TauSFT-Tau"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="bfloat16", device_map="auto")

The model is trained to emit tool calls in the tau-bench format; evaluate it inside a tau-bench environment rather than as a general-purpose chat model.

Notes

This is an intermediate RL checkpoint released for research purposes. It has not been safety-tuned, and it inherits the license and limitations of the underlying Qwen3-8B-Base model.