cs2881-hw1-combined

Combined RLAIF+RLVR — Setting B continued with GRPO on a within-batch 50/50 mix of the persona judge and the OpenBookQA verifier (per-type normalized reward). Both axes move (OpenBookQA +0.024, clf_style +0.257, register +1.105). Merged model at the repo root; the LoRA adapter is in adapter/.

Downloads last month
-
Safetensors
Model size
3B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for XujjHarvard/cs2881-hw1-combined

Base model

Qwen/Qwen2.5-3B
Finetuned
(1573)
this model