Omitted <think> at the start and almost 10k tokens to debug 2 JS functions
Hello,
I recently downloaded this model as a Q4_K_M GGUF in LMStudio and tested it on a little error finding problem in 2 JS functions.
As mentioned in the title, the model took almost 10k tokens to think and forgot the "" at the start.
2 questions:
- Is this amount of thinking tokens intended?
- Did the model also omit the "" for anyone else?
Thank you for your work!
Same, i just deleted it for this reason.
I said "Hi" and it's been responding for like 5 minutes because I forgot the <think>!
Hi everyone, we explicitly prefill the chat template with <think> to ensure it always generates the long chain-of-thought: https://huggingface.co/open-r1/OlympicCoder-7B/blob/a093c195a14f190b8228b12cf6cd180c21bfbeec/tokenizer_config.json#L198
In our experience, using sampling with temperature=0.6 and top_p=0.95 gives pretty coherent answers that don't get stuck in the infinite loop!