lm-sys / lm-sys/FastChat

Would an EvalPort-format export of MT-Bench be welcome?

Open
#3,926 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
39.5k
Forks
4.8k
PR merge metrics
No merged PRs in 30d

Description

Hi FastChat team — I maintain [EvalPort](https://github.com/adhabnr-ux/evalport), an open, framework-agnostic JSON spec for portable LLM eval datasets (Apache 2.0, `TestCase`/`Suite` schema with a real JSON Schema validator). It has a `benchmarks/` directory converting well-known public benchmarks into validated EvalPort suites, each with attribution and license preserved, so someone using a different harness can still run the same questions without a bespoke converter.

I read the actual data files rather than guessing at fields. `fastchat/llm_judge/data/mt_bench/question.jsonl`:

```json
{"question_id": 81, "category": "writing", "turns": ["Compose an engaging travel blog post about a recent trip to Hawaii...", "Rewrite your previous response. Start every sentence with the letter A."]}
```

plus `fastchat/llm_judge/data/mt_bench/reference_answer/gpt-4.jsonl` (reference answers keyed by `question_id`), and `fastchat/llm_judge/data/judge_prompts.jsonl` for the LLM-judge rubric.

This maps onto EvalPort cleanly: `turns` → a multi-turn `TestCase.input` list, `question_id`/`category` → `metadata`, the matching `reference_answer` entry → `expected_output` where present, and the judge prompt → an EvalPort `llm_judge` grader definition (MT-Bench's pairwise/single-answer grading is exactly the kind of grader EvalPort's spec has a slot for).

Would your maintainers be open to an EvalPort-format export of MT-Bench (and/or Vicuna-Bench, same structure) living in EvalPort's `benchmarks/` directory, with attribution to this repo/paper and the Apache-2.0 license preserved? I'd do the conversion work and validate it against EvalPort's real schema — filing this as a heads-up issue first rather than sending a PR out of nowhere. Also happy to link back to FastChat from EvalPort's benchmark listing either way. Let me know if this is of interest, or not a priority right now — no worries either way.

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reviewing fastchat/llm_judge/data/mt_bench/question.jsonl, the matching reference_answer/gpt-4.jsonl file, and judge_prompts.jsonl, then compare their contents with EvalPort's TestCase, Suite, and llm_judge schema. Done would require maintainer agreement on the proposed external export and a validated EvalPort benchmark with attribution and license information preserved.

Written by the indexing model from the issue text.

Assessment

Domain
data, machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.