How koii2k3/llama-oss works
LLaMA-OSS is a research workflow for distilling concise mathematical reasoning from GPT-OSS into a student model. Its documented path is teacher trace generation in multiple reasoning modes, answer verification and length-based curation, SFT, optional GRPO preparation and training, then benchmark evaluation. LLaMA-Factory and MS-SWIFT provide the training and inference frameworks. The README documents these stages, but the provided tree and excerpts do not expose the named workflow scripts or the curation implementation, so those links are documented rather than source-verified.
Main entry points: scripts Mode generation · main.py GRPO training · length_stats.py Length selection · evaluator.py Evaluation harness
Components 15
Trace Curation
- Mode generation
LLaMA-Factory/scripts - Answer verification
LLaMA-Factory/scripts - Length selection
LLaMA-Factory/length_stats.py
LLaMA-Factory
- SFT workflow
LLaMA-Factory/scripts - Chat inference
LLaMA-Factory/src/llamafactory/chat/hf_engine.py - SGLang engine
LLaMA-Factory/src/llamafactory/chat/sglang_engine.py - vLLM engine
LLaMA-Factory/src/llamafactory/chat/vllm_engine.py
MS-SWIFT
- GRPO data prep
- GRPO training
ms-swift/swift/cli/main.py - Inference engines
ms-swift/swift/llm/infer/infer_engine
Evaluation
- Benchmark evaluation
- Evaluation harness
LLaMA-Factory/lm-evaluation-harness/lm_eval/evaluator.py - GSM8K and MATH500
Other components
- Researcher
- GPT-OSS teacher
Connections 13
- Researcher → Mode generation: runs
- GPT-OSS teacher → Mode generation: generates traces
- Mode generation → Answer verification: checks answers
- Answer verification → Length selection: filters traces
- Length selection → SFT workflow: curated data
- Researcher → GRPO data prep: prepares data
- GRPO data prep → GRPO training: provides JSONL
- Researcher → GRPO training: starts training
- Researcher → Benchmark evaluation: runs evaluation
- Benchmark evaluation → Evaluation harness: evaluates models
- Evaluation harness → GSM8K and MATH500: reports scores
- SFT workflow → Chat inference: uses framework
- GRPO training → Inference engines: uses inference