finetune/generation_config.json

{
  "_comment": "Stop generation on <|im_end|> (248046). The merged model's config.json sets eos_token_id to <|endoftext|> (248044), which the chat format never emits at turn end, so generation ran to max_new_tokens. Upload this to the merged model repo with finetune/push_generation_config.py.",
  "bos_token_id": 248045,
  "pad_token_id": 248044,
  "eos_token_id": [248046, 248044],
  "do_sample": false,
  "max_new_tokens": 2048,
  "transformers_version": "5.7.0"
}