SR-judger

Paper: Blind Variation, Selective Retention: Improving Creative Idea Generation in AI Scientists

Code: https://github.com/zou-group/BVSR

Data: Xkev/BVSR-data

The selective-retention selector of BVSR: it scores a proposal (task specification + code) so that the lowest-ranked proposals can be discarded before verification.

Training

  • Base model: Qwen/Qwen3.5-9B-Base
  • Dataset: sr_data/train.jsonl.gz of Xkev/BVSR-data
  • Objective: pairwise loss between each run's top-5% verified proposals and the rest
  • Hyperparameters: lr=1e-5 (head 1e-3), cosine schedule with warm-up, epochs=1

License

Apache-2.0.

Downloads last month
10
Safetensors
Model size
8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Xkev/SR-judger

Finetuned
(659)
this model

Collection including Xkev/SR-judger

Paper for Xkev/SR-judger