ByteDance

/

Ouro-2.6B

like 66

Follow

ByteDance 4.31k

Problem

Both bos_token and eos_token are set to <|endoftext|> (id=0), but Ouro uses ChatML format where:

• bos_token should be <|im_start|> (id=1)
• eos_token should be <|im_end|> (id=2)

This causes issues with:

• Generation stopping: model never sees a proper EOS signal
• Tokenizer add_special_tokens: wrong BOS is prepended
• Downstream tools (vLLM, lm-eval-harness) that rely on eos_token_id for stop conditions

Fix

config.json:

• bos_token_id: 0 → 1
• eos_token_id: 0 → 2

tokenizer_config.json:

• bos_token: <|endoftext|> → <|im_start|>
• eos_token: <|endoftext|> → <|im_end|>

See also: same fix merged for the Thinking variant — https://huggingface.co/ByteDance/Ouro-1.4B-Thinking/discussions/4