aidenjhwu/ASearcher_en_no-math_Qwen3-8B-reject-sample

See our blog for details: Cut the Bill, Keep the Turns: Affordable Multi-Turn Search RL

This dataset is originally from https://huggingface.co/datasets/inclusionAI/ASearcher-train-data

We do filtering to original data:

1. 1.Remove Chinese samples
2. 2.Our wiki server does not handle Chinese retrieval well and may return garbled text;
3. 3.There are about 2k Chinese-related samples in the ASearcher dataset, which we remove entirely.
4. 4.Remove math problems
5. 5.Use formula patterns / specific regexes to filter out math questions.
6. 6.Reject sampling with n=8 on the pretrained model
7. 7.Use an initial search agent (based on the pretrained base model) to roll out 8 trajectories per sample;
8. 8.If all 8 trajectories have reward 0 or reward 1, discard the sample:
9. 9.all 0s: too difficult / annotation issues / the environment fails to retrieve useful information → training signal is too weak;
10. 10.all 1s: too easy / does not require multi-turn retrieval → wastes RL budget;