feat: visualising search traces

[DESU-CLUB]

Feature Request: Search Model Human Alignment Evaluation

Background

Jia Qi brought up an amazing paper today: [MixEval: Deriving Wisdom of the Crowd from LLM Benchmark Mixtures](https://proceedings.neurips.cc/paper_files/paper/2024/file/b1f34d7b4a03a3d80be8e72eb430dd81-Paper-Conference.pdf)

Gist of the paper: Compares a set of human Google searches with the distribution of different datasets.

Proposed Feature

Spin-off idea: Measure the closeness of search models' distribution with these human Google searches instead, so as to model how well our model aligns with human search preferences.

We can then also rank other models on this same dataset and make a qualitative evaluation on which model is the most "human-aligned" to search.

Implementation

• Adapt MixEval's methodology from LLM evaluation to search model alignment
• Use distributional similarity between human search patterns and search model outputs
• Create leaderboard ranking different search models on human alignment
• Build evaluation framework that's efficient and reproducible

Some open questions

• What dataset/where to collect the human google search data (Yuuki has been driving some collation of internal search data by sending his own search queries in #jan-model-internal/Random Question for Model from User on Discord (search for Random question from user for on discord to find the thread)
• Which dataset should be used to curate the distributions of different search models (At a high level this is open to discussion with the team on how to do this)

Expected Outcome

Standardized benchmark for measuring how well search models align with human search behavior and preferences.

A picture of how the evaluation could look like can be seen in this excalidraw.

[Yip-Jia-Qi]

I think one problem with this approach, is that we are assuming that if the search queries of Jan are human-like that is good for performance.

At a more fundamental level, i'm wondering if we are able to evaluate the diversity of search queries as an indicator of the model.

For even better, whether we can observe the path of the model through the embedding space as it is searching. If the queries are very focused in one area/clustered, thats a sign that its going into an infinite loop

[new5558]

Comments on the issue

1. Dataset for search queries

We can use LM Arena - search as human sourced search query to create search version of mix eval

2. Speed up eval process for Agentic LLM eval

We might need to convert E2E evaluation dataset into a process eval dataset, then eval each step indepentently to improve evaluation speed..

Observations:

• Most agentic search dataset are e2e dataset without golden process of the best path from calling tools to generating answer.

Open Questions:

• Can we use LLM judge to score each step of agentic search LLM?

Draft (WIP)
golden step
E2EA->B->C->D

1. A->?
2. A->B->?
3. A->B->C->?

we have to measure correlation of the

SimplQA, FreshQA, SealQA, Jan-exam-v1 -> multistep eval