Siye01/LLM-Fusion-Train
Multi-Domain RLVR Training Data Per-domain reinforcement-learning-with-verifiable-rewards (RLVR) training sets for five domains, plus the mixed-domain blend used as a joint-training baseline. Every subset uses the verl RLHF parquet schema: data_source, prompt, ability, reward_model, extra_info. Subsets Subset Rows Size Math 38,131 11.6 MiB Science 50,000 45.6 MiB Code 19,169 1,432.5 MiB IF 16,575 9.3 MiB Agent 10,229 0.2 MiB Mix 87,699 1,505.1… See the full description on the dataset page: https://huggingface.co/datasets/Siye01/LLM-Fusion-Train.
Multi-Domain RLVR Training Data
Per-domain reinforcement-learning-with-verifiable-rewards (RLVR) training sets for five domains, plus the mixed-domain blend used as a joint-training baseline. Every subset uses the verl RLHF parquet schema: data_source, prompt, ability, reward_model, extra_info.
Subsets
Mix is the exact blend used in our experiments: it subsamples Math and Science while using Code, IF, and Agent in full. It is released verbatim rather than as a sampling script, since re-running the sampler would not reproduce the same rows.
Provenance and licensing
Each subset is a reformatting of a public upstream dataset into verl's RLHF schema. Attribution and terms follow the upstream sources:
Code, IF, and Agent come from NVIDIA's Nemotron RL collection, released through NeMo Gym.
Usage
from datasets import load_dataset
math_train = load_dataset("<repo-id>", "Math", split="train")