These traces are released for non-commercial research on multimodal tool-use agents. They contain outputs of commercial models (Claude Opus 5.5, gpt-6-sol, Kimi K3), which remain subject to their providers' terms of use, including restrictions on using outputs to develop competing models. Images are not included; obtain them from MM-ToolSandbox under its license.

MM-ToolSandbox Agent Traces (s1)

This dataset holds agent runs on MM-ToolSandbox, Apple's multimodal tool-use benchmark: 30 scenarios × 3 repetitions per model. The models are three strong teachers (Claude Opus 5.5, gpt-6-sol and Kimi K3, still running) and one small open model (Qwen3.5-9B). Every run is graded by MM-ToolSandbox's judge.

The traces use an MMDU-style ShareGPT format, extended with tool calls, tool results and reasoning, so LLaMA-Factory can load them directly. Each scenario is also sorted into a bucket by comparing teacher and student. The intended use is studying where a small model falls short of a strong one, and how task-specific post-training might close the gap.

Results

Pass rate is the judge's task-completion verdict, averaged over 30 scenarios and 3 repetitions. A scenario counts as stable when it passes 3 of 3 runs, unstable at 1–2 of 3, and fail at 0 of 3.

Model	Pass rate	Passing runs	Stable / unstable / fail scenarios
Claude Opus 5.5	0.556 ± 0.017	50 / 90	14 / 4 / 12
gpt-6-sol	0.389 ± 0.017	35 / 90	7 / 10 / 13
Qwen3.5-9B	0.011 ± 0.017	1 / 90	0 / 1 / 29

| Kimi K3 | 0.472 (partial) | 25 / 53 | 5 / 5 / 7 |

Kimi K3 (in progress)

Kimi K3 is being added as a fourth model, with full reasoning (reasoning_content, see Record format). It runs on a subscription quota of a few runs per 5-hour window, so its files grow in batches and this section is regenerated with each upload. Its buckets against Qwen are in index/scenarios.csv (bucket_vs_kimik3).

Progress: 53 / 90 runs completed (last updated 2026-10-07 15:34 UTC). Buckets vs Qwen (scenarios with all 3 reps): 1=5, 2=0, 3=0, 4=1, other=11, pending=1 (bucket 1: scenario_0006, scenario_0017, scenario_0069, scenario_0075, scenario_0152).

The MM-ToolSandbox paper reports a success rate of about 0.19 for Qwen3.5-9B. Our setup has not yet been reconciled with the paper's, so treat the Qwen number as a lower bound for this harness, not as the model's true score.

Buckets

Each scenario is bucketed against each teacher. The columns bucket_vs_opus55 and bucket_vs_gpt6sol in index/scenarios.csv hold the result.

Bucket	Meaning	vs Opus 5.5	vs gpt-6-sol
1	Teacher stable, Qwen fails	13	7
2	Both stable	0	0
3	Teacher stable, Qwen unstable	1	0
4	Teacher not stable, Qwen passes at least once	0	1
other	Everything else	16	22

Bucket 1 vs Opus: scenario_0017, 0019, 0042, 0064, 0068, 0152, 0160, 0191, 0213, 0238, 0247, 0296, 0494.
Bucket 1 vs gpt-6-sol: scenario_0017, 0019, 0064, 0160, 0191, 0247, 0296.
Qwen's one pass is scenario_0028, repetition 2. That puts 0028 in bucket 3 against Opus and bucket 4 against gpt-6-sol.

index/buckets.md gives a one-line reason, quoted from the judge, for each scenario where Qwen fails. The common failure modes are:

looping until the 100-turn limit on long multi-tool tasks,
misreading values from images,
destructive or missed tool use.

Per-scenario results

Each mark is one run in repetition order: ✓ = judge pass, ✗ = fail, · = not run yet (Kimi K3 only). A model cell is bold when all three runs passed.

Scenario	Data / challenge / arrival	Task	Opus 5.5	gpt-6-sol	Kimi K3	Qwen3.5-9B	Bucket vs Opus	vs gpt-6-sol	vs Kimi
scenario_0006	docvqa / none / progressive	The user is an archivist working through a backlog of scanned historical documents. They…	✓✗✓	✗✗✗	✓✓✓	✗✗✗	other	other	1
scenario_0017	worldvqa / goal_change / late	The user is planning some upcoming purchases and wants to keep track of specific items th…	✓✓✓	✓✓✓	✓✓✓	✗✗✗	1	1	1
scenario_0019	chartqapro / state_mutation / progressive	Christina is preparing facts and outreach emails for an environmental awareness campaign,…	✓✓✓	✓✓✓	✓✓✗	✗✗✗	1	1	other
scenario_0028	docvqa / error_correction / mixed	Sierra is sorting through a batch of scanned historical business documents. She wants to…	✓✓✓	✗✗✗	✗✗✗	✗✗✓	3	4	4
scenario_0031	worldvqa / none / mixed	Hunter is planning a trip with his friend Charlie. They are using a Todoist project to tr…	✗✗✗	✗✗✗	✗✗✗	✗✗✗	other	other	other
scenario_0032	hiertext / error_correction / upfront	The user is planning a sponsorship pitch and needs to find the contact information for a…	✗✗✗	✗✗✓	✗✗✓	✗✗✗	other	other	other
scenario_0034	omnidocbench / none / mixed	Eric is organizing his personal notes, reading lists, and shopping tasks by extracting in…	✗✗✗	✗✗✗	✗✗✗	✗✗✗	other	other	other
scenario_0042	hiertext / error_correction / mixed	The user is participating in a city-wide scavenger hunt and trivia game. They are gatheri…	✓✓✓	✗✗✗	✓✗✓	✗✗✗	1	other	other
scenario_0064	worldvqa / none / upfront	The user is exploring various stage performances and wants to compile their soundtracks i…	✓✓✓	✓✓✓	✓✓✗	✗✗✗	1	1	other
scenario_0068	worldvqa / state_mutation / late	Ashley is setting up a new creative workspace. She has a fixed starting budget and is usi…	✓✓✓	✗✗✓	✗✓✗	✗✗✗	1	other	other
scenario_0069	screenspotpro / error_correction / mixed	The user is following a CMake tutorial to build OpenCV from source on a Linux virtual mac…	✓✓✗	✓✓✗	✓✓✓	✗✗✗	other	other	1
scenario_0075	screenspotpro / none / progressive	The user is working on a MATLAB project for 3D plot annotations and needs to validate fil…	✗✓✓	✓✗✗	✓✓✓	✗✗✗	other	other	1
scenario_0087	chartqapro / none / mixed	Christopher is compiling a report on specific health statistics from various historical a…	✗✗✗	✗✗✗	✗✗✗	✗✗✗	other	other	other
scenario_0092	docvqa / goal_change / upfront	Alexander is organizing a collection of historical documents, letters, and worksheets. He…	✗✗✗	✗✗✗	✗✗✗	✗✗✗	other	other	other
scenario_0102	hiertext / error_correction / upfront	Troy is documenting funny and interesting store signs he saw recently to share a compiled…	✗✗✗	✗✗✗	✗✗✗	✗✗✗	other	other	other
scenario_0134	infographicvqa / error_correction / progressive	Patricia is organizing events and campaigns for her local sports and health awareness clu…	✗✗✗	✗✗✗	✗✗✗	✗✗✗	other	other	other
scenario_0152	hiertext / goal_change / late	The user is planning a special dinner for a VIP guest and is using their local files and…	✓✓✓	✗✗✓	✓✓✓	✗✗✗	1	other	1
scenario_0160	chartqapro / goal_change / late	Sandra Martinez is gathering financial data across multiple assets and needs to compile a…	✓✓✓	✓✓✓	✓✓·	✗✗✗	1	1	pending
scenario_0181	worldvqa / state_mutation / progressive	Claudia is cataloging antique and vintage items she plans to sell at an upcoming auction.…	✗✗✗	✗✗✗	···	✗✗✗	other	other	no_data
scenario_0191	hiertext / none / progressive	The user is browsing various books and magazines and wants to purchase some of them on Am…	✓✓✓	✓✓✓	···	✗✗✗	1	1	no_data
scenario_0213	screenspotpro / goal_change / upfront	The user is researching trending AI models shown on Hugging Face in their Android emulato…	✓✓✓	✓✓✗	···	✗✗✗	1	other	no_data
scenario_0236	omnidocbench / error_correction / upfront	The user is digitizing a stack of old paper documents and business records, extracting co…	✗✗✗	✗✗✗	···	✗✗✗	other	other	no_data
scenario_0238	chartmuseum / error_correction / upfront	Dillon Wilson is analyzing several machine learning research charts and needs to filter h…	✓✓✓	✓✗✓	···	✗✗✗	1	other	no_data
scenario_0247	hiertext / state_mutation / mixed	The user is planning a new conference and is gathering inspiration from photos of past ev…	✓✓✓	✓✓✓	···	✗✗✗	1	1	no_data
scenario_0296	chartmuseum / none / progressive	Melissa is compiling various business, market, and demographic statistics from several re…	✓✓✓	✓✓✓	···	✗✗✗	1	1	no_data
scenario_0347	chartqapro / error_correction / upfront	The user is managing their weekly review process and wants to ensure all business divisio…	✓✓✗	✗✓✗	···	✗✗✗	other	other	no_data
scenario_0352	docvqa / state_mutation / mixed	The user is processing a batch of scanned documents, extracting contact information and c…	✗✗✗	✓✗✗	···	✗✗✗	other	other	no_data
scenario_0355	worldvqa / error_correction / upfront	The user is organizing a luxury showcase event and has collected images of specific furni…	✗✗✗	✗✗✗	···	✗✗✗	other	other	no_data
scenario_0494	screenspotpro / state_mutation / late	Vickie is setting up a workspace for evaluating language models. She is tracking her data…	✓✓✓	✗✓✓	···	✗✗✗	1	other	no_data
scenario_0552	infographicvqa / goal_change / upfront	The user is reviewing a collection of infographics to plan their content writing schedule…	✗✗✗	✗✗✗	···	✗✗✗	other	other	no_data
Totals		passes / runs	50 / 90	35 / 90	25 / 53	1 / 90

See index/runs.csv for per-run details (turns, tokens, wall time) and index/buckets.md for the judge's reasons.

Files

Path	Contents
`data/full/<model>.jsonl`	One record per run (90 per model): the whole trajectory
`data/per_step/<model>.jsonl`	One record per assistant step: the history up to that step, with that step's reasoning and action as the final turn. Opus 5.5 has 1,283 records, gpt-6-sol 1,979 and Qwen3.5-9B 2,304.
`index/scenarios.csv`	Per scenario: dataset type, challenge type, image arrival, each model's stable/unstable/fail status, and both buckets
`index/runs.csv`	Per run: model, repetition, pass, judge score, turns, tokens, API calls, wall time
`index/results.csv`	Full per-run results table from the sweep
`index/buckets.md`	Bucket tables with the judge's rationale
`index/scenarios.json`	The 30-scenario sample and its strata
`dataset_info.json`	LLaMA-Factory dataset registration for all six files

The per-step files exist because Qwen3.5's chat template keeps <think> only on the last assistant turn. In a full multi-turn record, every intermediate step's reasoning would be dropped at training time.

Record format

{
  "id": "scenario_0017_opus55_rep0",
  "image": ["worldvqa/images/1429.jpg", "..."],
  "system": "<agent system prompt>",
  "tools": [{"type": "function", "function": {"name": "...", "parameters": {}}}],
  "conversations": [
    {"from": "user", "value": "Image1: <ImageHere>. ... <request>"},
    {"from": "assistant", "value": "<think>...</think> <visible text>"},
    {"from": "function_call", "value": "{\"name\": \"...\", \"arguments\": {...}}"},
    {"from": "observation", "value": "<tool result>"}
  ],
  "meta": {"scenario": "scenario_0017", "model": "opus55", "rep": 0, "pass": true, "bucket": "1",
           "reasoning_source": "none", "sandbox_commit": "6397aba...", "n_steps": 12}
}

Envelope: matches MMDU (id, image, conversations, <ImageHere> placeholders). It adds LLaMA-Factory ShareGPT fields: system, tools, and function_call / observation turns. A reader that only understands plain MMDU will see just the dialogue.
Images: image holds paths relative to MM-ToolSandbox's image directory, in order of their <ImageHere> placeholders. The images are not redistributed here. Get them with MM-ToolSandbox, which draws on WorldVQA, HierText, ScreenSpot-Pro, DocVQA, ChartQAPro, OmniDocBench, InfographicVQA and ChartMuseum, and follow each source's license. For Qwen-style templates, map <ImageHere> to <|vision_start|><|image_pad|><|vision_end|>.
Reasoning: meta.reasoning_source records where the reasoning came from.
- none (Claude Opus 5.5): the benchmark's standard agent prompt was used, and the hidden thinking is not readable. Records carry only the visible text and tool calls, with no <think>.
- openai_summary (gpt-6-sol): the Responses API reasoning summary (summary: "detailed") appears as <think>. It is a summary, not the full chain of thought.
- qwen_native (Qwen3.5-9B): the model's own <think> text, split out by vLLM's qwen3 reasoning parser.
- kimi_native (Kimi K3, model id k3, reasoning effort max): the model's full reasoning_content for each turn, returned by Moonshot's API. As the API requires, the complete assistant message (including reasoning_content) is sent back on every later turn, so the chain of reasoning is continuous across tool calls.
Labels: meta.pass is the judge verdict for that run. meta.bucket is that teacher's bucket for the scenario; it is null for Qwen records, whose buckets are in index/scenarios.csv.

How the runs were produced

Benchmark: MM-ToolSandbox at commit 6397aba, using its nominal scenarios. The flags were --enable-tool-search --enable-coding-tool --image-input, with the user simulator GPT_5_4_2026_03_05_User and the judge Claude_4_5_Sonnet_Judge.
Sample: 30 of the 258 nominal scenarios. They are stratified over dataset type (all 8), challenge type (none 8, goal_change 6, state_mutation 6, error_correction 10) and image arrival (upfront 10, mixed 8, progressive 7, late 5).
Agents:
- Claude Opus 5.5: the shipped Claude_Opus_5_5_Agent, unmodified.
- gpt-6-sol: reasoning effort high, through the Responses API.
- Qwen3.5-9B: served by vLLM 0.29 with bf16 weights, fp8 KV cache, a 64k context, at most 16 images per request, and the qwen3_coder tool parser plus qwen3 reasoning parser.
Isolation: each run executed in its own process.
Infrastructure fixes, applied uniformly before the final runs:
- MM-ToolSandbox's code-execution memory cap was raised from 512 MB to 32 GB; the low cap caused spurious polars crashes.
- Qwen's context was raised from 12k to 64k, after 12k runs were cut off by the length limit.
- Every run affected by an infrastructure problem was re-run. The final data has no infrastructure errors.

Limitations

The sample is small: 30 scenarios and 3 repetitions, so one scenario moves a pass rate by about 0.033.
The Qwen3.5-9B result has not been reconciled with the paper's reported number (see the warning above).
Opus 5.5 traces contain no reasoning text. gpt-6-sol reasoning is a summary. Kimi K3 reasoning is the model's full reasoning text.
Kimi K3 is still running (see its progress line). Until it reaches 90 runs, its pass rate and buckets are partial.
The data inside tool results (people, emails, calendars) is synthetic AppWorld data.

License and terms

Traces: released for non-commercial research.
Model outputs: Claude Opus 5.5, gpt-6-sol and Kimi K3 outputs remain subject to Anthropic's, OpenAI's and Moonshot's terms of use.
Benchmark: MM-ToolSandbox is under the Apple License.
Images: not included. Follow each source dataset's license.

If you use this data, please also cite MM-ToolSandbox and MMDU.

Downloads last month: 69

Total file size:

416 MB