evalith 0.2.0

Neutral, local-first AI regression-testing tool (明镜 / Evalith)

明镜 / Evalith

English | 中文

Catch AI regressions before your users do.

A neutral, local-first AI regression-testing tool. Define a test set, run it against any model (DeepSeek / Qwen / OpenAI / Claude / …), score every case, and diff two runs to see exactly what got better or worse — then gate CI so a prompt or model change can't silently break your product.

Why Evalith

• Neutral & open. Your evaluation harness decides which model wins, so it shouldn't be owned by a model vendor. Evalith is vendor-independent and open source (Apache-2.0).
• Local-first. The core workflow runs entirely on your machine — no account, no upload, no network. Your prompts and test data stay with you.
• China models first-class. DeepSeek, Qwen and global models are first-class aliases (evalith models); a Chinese llm_judge ships in the box.
• Regressions, not vibes. diff and --fail-on-regression tell you which cases improved, regressed, or broke when you change a prompt, model, or version.

Install

Requires Python ≥ 3.10.

pip install -e .            # core: pydantic, pyyaml, typer (no network needed)
pip install -e ".[litellm]" # optional: real models (DeepSeek/Qwen/OpenAI/Claude/...)

Quickstart (offline, no API key)

# 1. Run the example eval — uses the offline `echo` model, passes 2/2
evalith run examples/eval.yaml

# 2. Tweak your prompt/model in examples/eval.yaml, then run again
evalith run examples/eval.yaml

# 3. List runs, then diff the two newest to spot regressions
evalith list
evalith diff <OLDER_RUN_ID> <NEWER_RUN_ID>

Gate CI on regressions

Fail a build when quality drops — two ways:

# Absolute gate: fail if fewer than 90% of checks pass (no baseline needed)
evalith run examples/eval.yaml --fail-under 0.9

# Relative gate: fail if any case regressed vs a baseline run
evalith diff <BASELINE_RUN_ID> <NEW_RUN_ID> --fail-on-regression

Both exit non-zero on failure, so CI stops the PR. This repo ships a composite GitHub Action — drop this into .github/workflows/eval.yml:

name: AI eval gate
on: [pull_request]
jobs:
  eval:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: dominciyue/Evalith_MingJing@main
        with:
          config: examples/eval.yaml
          fail-under: "0.9"

(See .github/workflows/eval-example.yml for a working copy using the offline demo.)

Shareable reports

Turn a run or a diff into Markdown (for PR comments) or a self-contained HTML page:

evalith report <RUN_ID> --format md                       # Markdown to stdout
evalith report <RUN_ID> --format html --output report.html # standalone HTML file
evalith diff <A> <B> --format md --output diff.md          # diff as Markdown

Reports include the pass rate, mean score, and — for real models — cost, token count, and latency.

Using real models (国产 first-class)

evalith models          # list first-class aliases + the env var each needs
export DEEPSEEK_API_KEY=sk-...
evalith run examples/eval.deepseek.yaml --concurrency 3
# -> Run <id> saved to .evalith/runs/<id>.json — 6/6 checks passed

Set model: to an alias (deepseek-chat, deepseek-reasoner, qwen-max, qwen-plus) or any LiteLLM id directly (gpt-4o-mini, claude-3-5-sonnet, …), then set that provider's API key. The llm_judge scorer can grade in Chinese with params: {language: zh}.

Scale

• --concurrency N runs cases in parallel (provider calls are I/O-bound), or set concurrency: in the config. Order of results is always preserved.
• Datasets load from YAML, JSON, CSV, or JSONL (examples/qa.jsonl).

Scorers

type	passes when
exact_match	output equals the case's expected
contains	output contains params.text (or the case's expected)
regex	output matches params.pattern
llm_judge	an LLM grades the output against params.criteria (params.language: en|zh)

How it works

run evaluates a config against a model and saves a Run — a JSON snapshot of every case's output, scores, tokens, cost, and latency — to .evalith/runs/. diff compares two saved runs case-by-case and labels each improved / regressed / unchanged / new / removed.

Status

v0.2 — single-turn prompt evaluation, file-based run store, run-to-run diff, CI gating (--fail-under, --fail-on-regression, GitHub Action), Markdown/HTML reports, concurrency, cost/token/latency tracking, and 国产 model aliases with a Chinese judge. Team/cloud features are on the roadmap. Issues and PRs welcome.

License

Apache-2.0. Copyright © 2026 Evalith (明镜) Authors.