OneReason-0.8B Frontier SFT Epoch 2 -> Reference-Free RLOO LoRA - Epoch 2
This repository contains one LoRA adapter checkpoint for the Kuaishou LLM4Rec competition.
Training
- Start: OneReason-0.8B competition base with the Frontier SFT Epoch 2 r64 adapter
- Method: G=16 live legal-SID sampling, five-tier SID reward, leave-one-out advantages without standard-deviation normalization, conditional GT-set anchor, no reference model, no KL, no GT injection, two epochs
- Frontier RLOO groups: 17,016 normalized groups x 2 epochs
- Completed epoch: 2
- LoRA rank / alpha / dropout: 64 / 64 / 0.0
- Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Files
adapter_model.safetensors: LoRA weightsadapter_config.json: the original training checkpoint config, preserved byte-for-byte
Evaluation status
Only a local Frontier-derived constrained beam-16 probe was run; no official competition score or performance improvement is claimed.
