nathon

fix: correct margin calculation in DPO training Remove redundant dpo_beta multiplication in margin computation. The chosen_rewards and rejected_rewards already include the dpo_beta factor, so margin should not multiply by dpo_beta...

Change "also releases" to "also released" to use correct past tense when describing already published research paper and data. Fixes #157

fix: #165 correct pig latin training examples for proper consonant cluster handling - Fix "split" → "it-splay" (move entire "spl" consonant cluster) - Fix "quantum" → "antum-quay" (treat "qu" as...