jvonrad/olmo-2-7b-finetranslation-wikifact-grpo-att-mlp-checkpoint

OLMo-2-7B FineTranslation + WIKI-FACT GRPO (Attention + MLP)

This model is based on allenai/OLMo-2-1124-7B.

It was first continuously pretrained on FineTranslations, and then GRPO-trained on WIKI-FACT, targeting attention and MLP layers.

Training recipe

1. Base model: allenai/OLMo-2-1124-7B
2. Continued pretraining on FineTranslations
3. GRPO training on WIKI-FACT
4. Trainable target modules: attention + MLP layers

Notes

• This repository contains the merged checkpoint.
• Intended for research on multilingual factuality and cross-lingual consistency.