OneReason-0.8B Frontier SFT Epoch 2 -> GRPO LoRA - Epoch 1
This repository contains the Epoch 1 LoRA adapter for the Kuaishou LLM4Rec competition.
Training
- Start adapter: official Frontier SFT Epoch 2
- Method: online GRPO with 8 live legal-SID candidates per prompt
- Data: 17,016 normalized Frontier recommendation groups
- Prompt rule:
/think and /no_think normalized to /no_think
- Reward target: final SID only; original thinking text is not used by the GRPO reward
- Forced GT: probability 0.5 when none of the 8 live candidates matches a GT
- LoRA rank / alpha / runtime dropout: 32 / 32 / 0.0
- Completed epoch: 1
Files
adapter_model.safetensors: LoRA weightsadapter_config.json: checkpoint configuration preserved byte-for-byte
The clean OneReason-0.8B competition base model is not included. Local probe and official competition evaluation are pending; no performance claim is made.
