youdotcom/minimax-m3-deepsearchqa-skill-eval

MiniMax M3 DeepSearchQA Skill Eval

Evaluates minimax/minimax-m3 on google/deepsearchqa using a Pi agent, You.com MCP tools, and a research skill optimized for this harness, model, and tool surface.

MiniMax M3 Medium Reasoning with the You.com research skill reached 73.17% adjusted F1 on DeepSearchQA, competitive with the paper's GPT-5 High Reasoning F1 result. Public artifacts are available for inspection and reproduction.

Links

• GitHub: https://github.com/youdotcom-oss/minimax-m3-deepsearchqa-skill-eval
• Hugging Face dataset: https://huggingface.co/datasets/youdotcom/minimax-m3-deepsearchqa-skill-eval

Project overview

This repository is an end-to-end evaluation project for minimax/minimax-m3 on Google's deepsearchqa dataset. It combines a research Skill, You.com MCP search/content tools, a Pi-based agent adapter, a generation pipeline, and grading/upload scripts.

The main flow is:

DeepSearchQA dataset
        │
        ▼
scaffold ──► prompts.jsonl ──► generate ──► trajectories.jsonl
                                                │
                                                ▼
                                      grade ──► graded.jsonl + summary.json
                                                │
                                                ▼
                                      export/upload ──► public artifacts

Repository structure

.
├── src/                              Core adapter, grader, MCP/session integration
├── scripts/                          Scaffold, generate, grade, export, download, query, upload
├── datasets/
│   ├── public/                       Public dataset provenance and documentation
│   └── user/                         User-supplied input format and safe example
├── integrations/minimax-code/
│   ├── README.md                     MiniMax Code setup and smoke-test guide
│   ├── plugin/                       Portable you-web Plugin package
│   └── smoke-prompts/                Search, contents, and installed-Plugin test prompts
├── .minimax/skills/you-web/          Project-level MiniMax Code Skill
├── data/                             Local generated artifacts; not committed
├── package.json                      Bun scripts and dependencies
└── README.md                         Project overview, results, reproduction, and publishing

The project-level Skill and portable Plugin contain the same you-web workflow. The former is loaded when MiniMax Code opens this repository; the latter can be installed and used from any working directory.

Quick start

Install dependencies, configure the required credentials locally, and run a one-question smoke:

bun install
OPENROUTER_API_KEY=... YDC_API_KEY=... FORCE=1 LIMIT=1 K=1 CONCURRENCY=1 bun run eval
cat data/summary.json

Use bun run check for local code checks. See Reproduce for the full evaluation and MiniMax Code integration for the Skill/Plugin workflow.

Results

Full run completed on 2026-08-21 with minimax/minimax-m3 at THINKING_LEVEL=medium, the You.com MCP you-search and you-contents tools, and a MiniMax-oriented research skill optimized via an auto-research loop over the model, tool surface, and harness combination.

These are eval results and artifact facts, not a paper claim. The uploaded prompts.jsonl, results.jsonl, trajectories.jsonl, graded.jsonl, and summary.json are the source of record.

Metric	Value
Adjusted F1 score	73.17%
Adjusted pass rate (score >= 0.8)	66.67%
Adjusted exact pass@3 (score >= 0.8)	83.93%
Fully correct (score = 1.0)	56.58%
Trials	2,700
Tasks	900
Total cost	$478.41
P50 / P95 latency	54.3s / 238.8s

Summary metrics

Metric	Raw	Adjusted
Trials	2,700	2,688
Tasks	900	896
Average F1 score	72.84%	73.17%
Pass rate (score >= 0.8)	66.37%	66.67%
Exact pass@3 (score >= 0.8)	83.56%	83.93%
Ungradable trials	12	-
Ungradable tasks	4	-

Cost, latency, and tool use

Cost totals:

Metric	Value
Total cost	$478.41
Model cost	$328.14
You.com API cost	$150.27
Average cost per task	$0.532

Per-trial distributions:

Metric	Average	P50	P90	P95
Cost	$0.177	$0.095	$0.431	$0.606
End-to-end latency	82.5s	54.3s	177.7s	238.8s
Tool calls	28.62	20	60	78

Tool call totals:

Metric	Value
Total tool calls	77,286
Search calls	24,999
Contents calls	13,641

Benchmark context

The DeepSearchQA paper reports single-response F1 plus categorical rates such as Fully Correct and Fully Incorrect. The project summary's passRate and exactPassAtK are useful operational metrics, but they are not the same as the paper's Fully Correct rate because this repo's pass threshold is score >= 0.8.

Using the graded rows to compute paper-style adjusted trial metrics:

Metric	Value
F1	73.17%
Fully correct (score = 1.0)	56.58%
Fully incorrect (score = 0.0)	19.01%
Correct with extraneous answers	11.31%
Partially correct	13.10%

For context, the DeepSearchQA paper's Table 4 reports GPT-5 High Reasoning at 73.24 F1, Gemini 3 Pro Preview at 76.86 F1, GPT-5 Pro High Reasoning at 78.98 F1, and Gemini Deep Research Agent at 81.90 F1. This run is therefore best described as competitive with the paper's GPT-5 High Reasoning F1 result, while still behind the top Deep Research agents and with a higher fully-incorrect rate than the leaders.

Eval setup

• Dataset: google/deepsearchqa
• Model: minimax/minimax-m3
• K: 3
• Generation concurrency: 3
• Thinking level: medium
• Tools: You.com MCP you-search and you-contents
• Judge: deepseek/deepseek-v4-flash-0731
• Judge fallback: qwen/qwen3.6-flash
• Adapter stdout cap: 50_000_000 bytes per trial

Evidence artifacts

• prompts.jsonl — DeepSearchQA prompts in agent-eval-harness task format.
• results.jsonl — flat, viewer-safe per-trial result rows derived from graded.jsonl.
• trajectories.jsonl — raw agent-eval-harness run output.
• graded.jsonl — raw agent-eval-harness grade output.
• summary.json — aggregate raw and adjusted metrics.

Generated files live under local data/ and are uploaded to the Hugging Face dataset root. They are not committed to GitHub.

The uploaded Hugging Face dataset card configures the Dataset Viewer to load stable, flat results.jsonl rows and prompts.jsonl prompts. The raw graded.jsonl and trajectories.jsonl harness artifacts remain downloadable, but are intentionally excluded from the viewer because they contain heterogeneous nested records.

Inspect published results

To inspect the public results without rerunning the full eval, download the published artifacts into data/:

bun run download -- --dry-run
bun run download -- --yes

By default this downloads summary.json, prompts.jsonl, results.jsonl, and graded.jsonl. Use --all to include the much larger trajectories.jsonl:

bun run download -- --all --yes

For ad-hoc questions over large data/*.jsonl artifacts, use the ClickHouse helper instead of loading JSONL into memory:

bun run download -- --yes
bun run query -- --list
bun run query -- summary --dry-run
bun run query -- summary
bun run query -- failures
bun run query -- cost-outliers
bun run query -- latency-outliers

This requires clickhouse-local on PATH. If your ClickHouse binary is invoked another way, set CLICKHOUSE_LOCAL, for example:

CLICKHOUSE_LOCAL="./clickhouse local" bun run query -- summary

Reproduce

Run with Bun and uv installed:

bun install
bun run eval
bun run export-results
bun run upload

uv is used by bun run upload and bun run download to run the Python Hugging Face scripts and install their Python dependencies.

Required environment variables:

OPENROUTER_API_KEY=...
YDC_API_KEY=...

Start with a one-question smoke run before launching the full eval:

FORCE=1 LIMIT=1 K=1 CONCURRENCY=1 bun run eval
cat data/summary.json

FORCE=1 clears prior generated artifacts for the command, which is useful when rerunning a smoke after code changes. Without FORCE=1, generate and grade resume from existing data/ files.

The smoke run validates plumbing. A score of 0 can still be a valid smoke result if the model missed the answer. Check that process.totalToolCalls is greater than 0 and that data/graded.jsonl does not show adapter_invalid_result.

If the smoke succeeds, run the full local eval:

caffeinate -dimsu bun run eval

Then upload the generated artifacts:

huggingface-cli login
bun run export-results
bun run upload

Commands

bun run scaffold    # DeepSearchQA -> data/prompts.jsonl
bun run generate    # prompts.jsonl -> trajectories.jsonl
bun run grade       # trajectories.jsonl -> graded.jsonl + summary.json
bun run eval        # scaffold + generate + grade
bun run export-results # graded.jsonl -> results.jsonl
bun run download    # download published artifacts from HF into data/
bun run upload      # upload README.md and data/* to HF
bun run query       # query large JSONL artifacts with clickhouse-local
bun run check       # typecheck + tests

Metrics definitions

summary.json reports raw metrics over all rows and adjusted metrics excluding ungradable rows with missing gold answers. exactPassAtK is task-level: a task passes if any of its K trials scores at least 0.8.

summary.json also reports cost:

• modelCostUsd from OpenRouter usage metadata.
• youApiCostUsd estimated from You.com pricing, currently $5/1k Search calls and $1/1k full-page extraction or Contents pages.
• searchExtractionPages and searchExtractionCostUsd for Search extraction_mode: "full_page" pages, counted across web and news results.
• totalCostUsd as model plus You.com API cost.

latency.averageEndToEndMs reports average harness-measured wall time from adapter invocation to final adapter output, using trial.invocation.durationMs.

Publishing to Hugging Face

Optional upload settings:

HF_DATASET_REPO=<your-hf-namespace>/<your-dataset-repo>
HF_REVISION=main

The uploader runs through uv and uses the official Python huggingface_hub client with hf_xet, which streams large files and can resume interrupted uploads. Full artifact uploads require a current data/results.jsonl; run bun run export-results after grading and before upload. The uploader uses your huggingface-cli login token, or HF_TOKEN if that environment variable is set. For maximum throughput on large artifacts, optionally set:

HF_XET_HIGH_PERFORMANCE=1

To update only the Hugging Face dataset card, without scanning the large artifacts:

bun run upload -- --card-only

The uploader prepends Hugging Face dataset-card metadata to the uploaded README.md; the GitHub README intentionally omits that YAML front matter. The uploaded dataset card configures the Hugging Face Dataset Viewer to use results.jsonl and prompts.jsonl, while leaving the raw graded.jsonl and trajectories.jsonl artifacts downloadable.

License

MIT. The generated Hugging Face dataset card declares license: mit.

MiniMax Code integration

This repository also provides a project Skill for local MiniMax Code at .minimax/skills/you-web/SKILL.md. It reuses the focused You.com workflow for you-search and you-contents.

See the MiniMax Code integration guide for project-level MCP setup, and the portable Plugin guide for installation from an arbitrary directory. The reproducible installed-Plugin test is in plugin-acceptance.md.

The MCP server used by this project is:

https://api.you.com/mcp?tools=you-search,you-contents

Keep authentication in local environment or local MCP configuration. Do not commit API keys, customer data, or private evaluation cases.