evalpair MCP server

evalpair

Run an LLM eval suite against a model, cache every call, and compare two runs item by item: did the new model, prompt or quantization actually change anything, or is the difference noise?

Most eval tools print two scores side by side. That reads a 2-point gap on 160 questions as a result, when it usually isn't one. evalpair pairs the two runs by item, puts a bootstrap interval on the difference, counts the items that got worse and better, and can measure a noise floor: the same model asked the same questions twice. In my own studies the same prompt sent twice at temperature 0 changed up to 3% of answers, so an answer-change rate only means something above that.

It's a CLI and an MCP server, so an agent can run an eval and read the verdict itself. It only calls the model server you point it at, and keeps everything in one local folder.

Quick start

uvx evalpair suites
uvx evalpair run gsm8k-160 -m llama3.1:8b --repeat --label llama
uvx evalpair run gsm8k-160 -m qwen3:8b --think off --label qwen3
uvx evalpair compare llama qwen3

run prints the score with a 95% interval and the first failures, unparsed and truncated replies first (those are usually the parser's or the token budget's fault, not the model's). compare prints both scores, the paired difference with its interval, how many items got worse and better, the noise floor, and ends with a plain-language verdict: whether the score difference is distinguishable from zero, and whether the answers moved more than the same model asked twice.

A real comparison: Ollama's default 4-bit Llama 3.1 8B against its 2-bit build, on an M5 Pro laptop (--num-ctx 4096; about 40 minutes for both runs):

$ evalpair compare llama-q4 llama-q2
A            llama-q4: 84.4% [79.4%, 90.0%]
B            llama-q2: 60.6% [53.1%, 68.1%]
B − A        -23.8 pp [-31.9 pp, -16.2 pp]
changed      40.6% [33.1%, 48.1%] of 160 items (43 worse in B, 5 better)
floor (A)    0.0% [0.0%, 0.0%]; excess +40.6 pp [+33.1 pp, +48.1 pp]

B scores lower than A: -23.8 pp [-31.9 pp, -16.2 pp], paired over 160 items. Answers changed on 40.6% of
items (43 got worse, 5 better). The noise floor (the same model asked twice) is 0.0% for A, so the change
is beyond noise.
  worse in B:  gsm8k-0259  '168' → '52'
  ...

Both scores are the ones the quantization study measured for these builds on the same items.

Interrupt a run and run it again: every reply is cached, so it picks up where it stopped.

As an MCP server

claude mcp add evalpair -- uvx evalpair mcp

or in any MCP client's config: {"command": "uvx", "args": ["evalpair", "mcp"]}.

Tool	What it does
list_suites	Suites you can run: name, item count, checker, what it measures
run_eval	Run a suite against one model and save the run (calls the model; cached replies are reused)
list_runs	Saved runs, newest first
show_run	A run's score with its interval, its noise floor, and the end of failed replies
compare_runs	Two runs paired by item: difference with an interval, flips each way, change vs. noise floor, examples

Everything except run_eval only reads saved runs.

Backends

• Ollama (default): --backend ollama, server at http://localhost:11434. The run records the model's digest, so you can tell later exactly which build was tested. --think on|off for models that think, --num-ctx for the context window.
• Any OpenAI-compatible server: --backend openai --base-url http://localhost:8000/v1 for vLLM, llama.cpp's server or LM Studio, or a hosted API. A bearer token is read from EVALPAIR_API_KEY (or OPENAI_API_KEY) and never written to disk.

Temperature defaults to 0 and the seed to 7. Both are part of the cache key, along with the server URL, model, messages, token budget and thinking setting.

Suites

Three are bundled. They're the item sets from my Measuring LLM Systems studies, with those studies' prompts word for word, so a run can be checked against the published numbers:

Suite	Items	Checker	Source
gsm8k-160	160	number	GSM8K test (MIT)
mmlu-hs-160	160	choice, stratified by subject	MMLU, four high-school subjects (MIT)
e2e-extract-150	150	json_fields	E2E NLG restaurant records (CC BY-SA 4.0)

To make your own, create a folder with two files, in ~/.evalpair/suites/<name>/ or anywhere you then pass by path:

# suite.toml
description = "Capitals, answered in one word"
checker = "exact"
max_tokens = 64
system = "Answer in one word."           # optional
user = "What is the capital of {country}?"
group = "region"                         # optional: stratify the bootstrap by this item field

{"id": "fr", "country": "France", "region": "Europe", "gold": "Paris"}
{"id": "pe", "country": "Peru", "region": "Americas", "gold": ["Lima", "Ciudad de los Reyes"]}

{field} in a prompt is filled from the item; a list of strings becomes lettered options (A. …), anything else is shown as indented JSON. Use {{ and }} for literal braces.

Checker	Scores	Gold
number	the number after "Answer:", else \boxed{}, else the last number	a number
choice	the letter after "Answer:", with fallbacks (letters set by checker_options.letters)	a letter
exact	the text after "Answer:", or the last line, normalised	a string or list of strings
contains	whether a gold string appears as whole words	a string or list
regex	the last match of checker_options.pattern (first group)	a string or list
json_fields	the share of the gold object's fields the reply's JSON gets right (null must be null)	an object
chrf	character n-gram F-score (sacreBLEU's chrF) against a reference text, for free text such as commit messages or summaries; checker_options.line = "first" compares first lines only	a string or list

Checkers are deterministic code: no LLM grades another LLM here. The number, choice and json_fields parsers come from the studies, and re-scoring all 9,870 replies from the quantization study's main run with them gives the study's own scores on every one (scripts/check_against_studies.py).

How the numbers are computed

• Unit of resampling: the item. 2,000 bootstrap draws, percentile intervals, seeded, so the same runs always give the same interval. With a group, items are resampled within each group.
• Comparisons are paired over the items both runs answered, using the first pass of each. If the two runs used different versions of a suite (the hash of its spec and items differs), the comparison says so.
• Noise floor: with --repeat, every item is sent a second time as a separate call (its own cache entry). The floor is the share of items whose parsed answer changed between the passes. A comparison calls a change "beyond noise" only if, per item, (changed between runs) minus (changed between the passes) has an interval above zero for every run that has a floor.
• Items are taken in file order, so --limit 50 is the same 50 items every time and limited runs still pair up.

Where things live

$EVALPAIR_HOME, or ~/.evalpair:

cache.sqlite           every reply, keyed by request
runs/<run>/meta.json   suite and its hash, model, digest, backend, options, counts, timings
runs/<run>/rows.jsonl  one line per item and pass: reply, parsed answer, score, tokens
suites/<name>/         your suites

Limits

• One model per run, chat-style prompts, one turn. No tool use, multi-turn or constrained decoding yet.
• Deterministic checkers only, so open-ended answers need a suite whose answer can be checked by code.
• The noise floor measures run-to-run variation on the same server. It doesn't cover a different machine or a server update.

License

MIT for the code. The bundled suites keep their sources' licenses (MIT for GSM8K and MMLU, CC BY-SA 4.0 for E2E NLG); each suite.toml names its license and source.