RationalPursuit/Qwen3-4B-Stratos-R1-GPT-OSS-120B-Blend-SFT

Model Card

Experimental research artifact. Base: `Qwen/Qwen3-4B-Instruct-2507`. SFT on `bespokelabs/Bespoke-Stratos-17k` prompts with a 50/50 blend of two teachers: each of the 14,768 paired prompts carries the trace of exactly one teacher, 7,384 from DeepSeek-R1 and 7,384 from gpt-oss-120b (reasoning effort medium, temperature 1.0), balanced per domain.

Same base, LoRA recipe, prompt set, and 24,576-token cap as the single-lineage students in this collection, varying only the teacher mix. Trained one epoch; validation loss on held-out prompts at the end of training was 0.423 on R1 traces and 0.619 on gpt-oss traces.

Eval (Stratos held-out, 508 prompts): see the collection's results once the held-out generation is recorded.

Out-of-Scope Use

Research artifact only — not intended for production use.

Framework versions

• —PEFT 0.19.1
• —TRL 1.8.0
• —Transformers 5.13.0