-
v10: bump README to retrigger CI eval
-
v10: SFT + Belebele + DPO 100% hard negatives
-
v9: bump README to retrigger CI eval
-
v9: SFT(MMLU+Belebele) + DPO with hard negatives (37%)
-
v8: update README — retrigger CI eval
-
v8: SFT (Global-MMLU + Belebele + INCLUDE regional knowledge) + DPO, fp16 merged
-
v7: SFT (volume 6000/lang Global-MMLU + Belebele, CoT answer-conditioned) + DPO, fp16 merged
-
v6: SFT (Global-MMLU+Belebele CoT) + DPO on preference pairs, fp16 merged
-
v5: SFT LoRA + Global-MMLU CoT + Belebele (5 langs), 1 epoch, fp16 merged
-
v4: SFT LoRA + answer-conditioned CoT (real reasoning), 1 epoch, fp16 merged
-
Clean re-push v3: remove stale sharded weights + index
-
Add model card (retrigger v3 eval)
-
v3: SFT LoRA + English CoT template (assistant: 'The answer is X, which corresponds to option Y. \boxed{Y}'), 1 epoch, fp16 merged
-
v2: SFT LoRA + format instruction in user prompt (5 langs), 2 epochs, fp16 merged
-
v1: SFT LoRA on Global-MMLU 5 langs (it/es/zh/ru/hi), 13.5k samples, 1 epoch, fp16 merged
-
v0: Qwen3-1.7B base + chat_template thinking OFF + system prompt boxed + gen_config tuned
-
initial commit
