Sakana AIDiscoPOP
DiscoPOP-zephyr-7b-gemma
Sakana fine-tuned Zephyr's Gemma supervised checkpoint on preference pairs using DiscoPOP instead of ordinary DPO.
Research release
The preference objective is the experimental change
Training uses argilla/dpo-mix-7k. The card says the setup matches Zephyr's preference-training recipe apart from replacing DPO with DiscoPOP. The published run used two training epochs.DiscoPOP-zephyr-7b-gemma official model card
Working with this release
Practical guidance. For a method comparison, hold prompts and decoding settings constant. Attribute results to this trained checkpoint rather than to all Gemma models.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | SakanaAI/DiscoPOP-zephyr-7b-gemma |
| Ancestry | HuggingFaceH4/zephyr-7b-gemma-sft-v0.1.DiscoPOP-zephyr-7b-gemma official model card |
| Terms | Gemma terms.DiscoPOP-zephyr-7b-gemma official model card |