Explain LLM predictions, using logprobs to highlight token probability

[lopuhin]

This can allow to see when model is less certain:

TODO:

• tests
• tutorial, including limitations in case of CoT
• mention as an extra method in README and docs

[lopuhin]

Would be cool to support that with open source models, likely in a future PR, I had a look:

• ollama does not support logprobs, but the features is being discussed Provide logits or logprobs in the API ollama/ollama#2415
• vllm does seem to support that but it's probably not the most popular hobbyist option
• mlx-llm also supports that, probably makes sense to use in a tutorial.

[Copilot]

Copilot reviewed 10 out of 18 changed files in this pull request and generated 2 comments.

Files not reviewed (8)

• README.rst: Language not supported
• docs/source/libraries/index.rst: Language not supported
• docs/source/libraries/openai.rst: Language not supported
• docs/source/overview.rst: Language not supported
• docs/source/tutorials/explain_llm_logprobs.rst: Language not supported
• docs/update-notebooks.sh: Language not supported
• eli5/templates/weighted_spans.html: Language not supported
• tox.ini: Language not supported

[codelion]

You can use optillm for that - https://github.com/codelion/optillm see algorithmicsuperintelligence/optillm#168 (reply in thread) you can also visualize the ouputs using the logprobs visualizer - https://huggingface.co/spaces/codelion/LogProbsVisualizer

[lopuhin]

Thanks you @codelion , do you mind if I add those links to the end of the tutorial here https://eli5.readthedocs.io/en/stable/tutorials/explain_llm_logprobs.html#open-source-and-other-models ?

[codelion]

@lopuhin please go ahead, thanks!