Draft notes and calculation #4
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
willpoulett
commented
Dec 27, 2024
|
|
||
| An alternative benchmark is presented by the paper [Evaluating Large Language Models at Evaluating Instruction Following](https://arxiv.org/pdf/2310.07641) which creates the LLM-Bar benchmark. This benchmark focusses on objective performance beyond correctness through both natural and adversarial human annotation. The work also includes a discussion of prompting strategies and could be useful for considering any prompt instructions are being assessed in terms of preferencing. | ||
|
|
||
| MT-bench is a third benchmark described in [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://arxiv.org/pdf/2306.05685) **DN: More on this** |
There was a problem hiding this comment.
Does this need adding to?
willpoulett
commented
Dec 27, 2024
| Overall, all of these benchmarks highlight limitations and are not recommended for specific use-cases but only to support initial development stages and to provide some additional considerations when choosing a model for the judge. | ||
|
|
||
| ## 2. Metrics to assess against human evaluations | ||
| Alignment metrics are used to compare the evaluations from human judges against LLM judges. These look at the amount of overlap **DN: Sentence of alignement metrics** |
There was a problem hiding this comment.
Does this also need adding to?
willpoulett
commented
Dec 27, 2024
| ## 3. Identify strategies to improve the judge | ||
| [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://arxiv.org/pdf/2306.05685) & [JudgeLM: Fine-tuned large language models are scalable judges](https://arxiv.org/pdf/2310.17631) identify a range of potential biases that judges can exhibit. These include: | ||
|
|
||
| - Position bias (**DN: MORE** |
There was a problem hiding this comment.
Positional and Knowledge bias need a quick sentence
willpoulett
commented
Dec 27, 2024
| These three mitigations are shown to enhance the judge of interest. | ||
|
|
||
| ## Additional Considerations | ||
| [When combinations of humans and AI are useful: A systematic review and meta-analysis](https://www.nature.com/articles/s41562-024-02024-1) finds an interesting conclusion from a systematic review of 106 studies for performance of evaluations by humans-alone vs AI alone vs human-AI combinations. It summarised that Human-ai performed worse than the best of human alone or ai alone, but when humans alone outperformed AI alone, they found performance gains in the combination but when AI outperformed humans alone, they found losses. The also found a trend of negative performance for human-ai synergy for classification tasks whilst a positive uplift in performance for creative tasks. |
There was a problem hiding this comment.
I found this a little confusing:
but when humans alone outperformed AI alone, they found performance gains in the combination but when AI outperformed humans alone, they found losses.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.