CoolFace
Datasetpublic

yavuz-ai/self-reward-collapse-terse

self-reward-collapse-terse Per-round trajectory and held-out samples from an iterative self-training loop on GSM8K (Qwen2.5-7B-Instruct, LoRA DPO, 6 rounds). Same loop every arm; only the preference-label source differs. This dataset is the terse arm: pairs labelled by a deliberately gameable brevity reward (the shortest of K samples wins). Study question: does a model training on its own judgment collapse? Short answer on verifiable math: the reward can be hacked, but… See the full description on the dataset page: https://huggingface.co/datasets/yavuz-ai/self-reward-collapse-terse.

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face
0likes82downloads

yavuz-ai/self-reward-collapse-terse · main · files are served by the source, never re-hosted here