willamazon1/sdft-tau-lora-iter240
The willamazon1/sdft-tau-lora-iter240 is an 8 billion parameter Qwen3-Base model, enhanced through a multi-stage Supervised Fine-Tuning (SFT) process and further optimized with a merged tau-bench Reinforcement Learning (RL) LoRA adapter. This model is designed for specific agentic tasks, particularly within retail contexts, leveraging its specialized training on the tau-bench dataset. It offers full weights, ready for direct use with standard inference libraries without requiring PEFT.
Loading preview...
Overview
The willamazon1/sdft-tau-lora-iter240 is an 8 billion parameter model based on Qwen/Qwen3-8B-Base. It has undergone a multi-stage Supervised Fine-Tuning (SFT) process, referred to as SDFT, which includes training on Math, Sea, Search, TauSFT, and Tau-IF datasets. A significant differentiator is the integration of a tau-bench RL LoRA adapter, merged directly into the full weights, making it immediately usable without PEFT.
This iteration, 240, represents 80 additional RL steps compared to its predecessor, willamazon1/sdft-tau-lora-iter160. The RL training utilized GSPO on the tau-bench retail dataset's training split, with a LoRA-only configuration targeting linear_qkv, linear_proj, linear_fc1, and linear_fc2 modules across all 36 layers.
Key Capabilities
- Specialized Agentic Performance: Optimized for tasks within the retail domain due to its specific tau-bench RL training.
- Direct Deployment: Provided as full, merged weights, allowing for straightforward loading with
transformers, SGLang, or vLLM. - Enhanced from Base Model: Builds upon the robust Qwen3-8B-Base architecture with extensive SFT and RL.
Usage Considerations
- Prompt Format: Users should adhere to the prompt format used by the tau-bench agent during training, as this model is derived from a base (non-instruct) Qwen3 checkpoint plus SFT/RL stages, not a generic chat template.
- LoRA Configuration: The LoRA adapter was trained with a rank of 16 and alpha of 32, applying a scaling of 2.0, and targeting 144 modules across 36 layers. Verification confirmed the adapter's active training and significant delta size.