# pulze-intent-v0.1

Intent-tuned LLM router that selects the best LLM for a user query. Use with [knn-router](https://github.com/pulzeai-oss/knn-router).

## Models

- claude-3-haiku-20240307
- claude-3-opus-20240229
- claude-3-sonnet-20240229
- command-r
- command-r-plus
- dbrx-instruct
- gpt-3.5-turbo-0125
- gpt-4-turbo-2024-04-09
- llama-3-70b-instruct
- mistral-large
- mistral-medium
- mistral-small
- mixtral-8x7b-instruct

## Data

### Prompts and Intent Categories
Prompt and intent categories are derived from the [GAIR-NLP/Auto-J scenario classification dataset](https://github.com/GAIR-NLP/auto-j/blob/2ae17a3965d933232e9cd50302aa0f176249c83b/README.md?plain=1#L582).

Citation:
```
@article{li2023generative,
  title={Generative Judge for Evaluating Alignment},
  author={Li, Junlong and Sun, Shichao and Yuan, Weizhe and Fan, Run-Ze and Zhao, Hai and Liu, Pengfei},
  journal={arXiv preprint arXiv:2310.05470},
  year={2023}
}
```

### Response Evaluation

Candidate model responses were evaluated pairwise using `openai/gpt-4-turbo-2024-04-09`, with the following prompt:

```
You are an expert, impartial judge tasked with evaluating the quality of responses generated by two AI assistants.

Think step by step, and evaluate the responses, <response1> and <response2> to the instruction, <instruction>. Follow these guidelines:
- Avoid any position bias and ensure that the order in which the responses were presented does not influence your judgement
- Do not allow the length of the responses to influence your judgement - a concise response can be as effective as a longer one
- Consider factors such as adherence to the given instruction, helpfulness, relevance, accuracy, depth, creativity, and level of detail
- Be as objective as possible

Make your decision on which of the two responses is better for the given instruction from the following choices:
If <response1> is better, use "1".
If <response2> is better, use "2".
If both answers are equally good, use "0".
If both answers are equally bad, use "0".

<instruction>
{INSTRUCTION}
</instruction>

<response1>
{RESPONSE1}
</response1>

<response2>
{RESPONSE2}
</response2>
```

Each pair of models is subject to 2 matches, with the positions of the respective responses swapped in the evaluation prompt. A model is considered a winner only if 
it wins both matches.

For each prompt, we then compute Bradley-Terry scores for the respective models using the same [method](https://github.com/lm-sys/FastChat/blob/f2e6ca964af7ad0585cadcf16ab98e57297e2133/fastchat/serve/monitor/elo_analysis.py#L57) as that used in the [LMSYS Chatbot Arena Leaderboard](https://chat.lmsys.org/?leaderboard). Finally, we normalize all scores to a scale from 0 to 1 for interoperability with other weighted ranking systems.

## Model

The embedding model was generated by first fine-tuning [`BAAI/bge-base-en-v1.5`](https://huggingface.co/BAAI/bge-base-en-v1.5) with the intent categories from the dataset above, using contrastive learning with cosine similarity loss, and subsequently merging the resultant model with the base model at a 3:2 ratio.