AutoDataBench: DeepSeek V4 Pro Rollouts

Project · Blog · Code · Download

This dataset contains 144 DeepSeek V4 Pro attempts on the 24 original tasks used by AutoDataBench, together with verifier results and 24 behavioral rubrics derived from the attempts. Each task has six independent attempts, including both successes and failures.

AutoDataBench studies whether agents can produce useful executable tasks for agent training. An author agent receives an original task and the target model's attempts, then builds a new task aimed at the same behavioral challenges. The delivered task is evaluated for validity and novelty, suitable difficulty, and coverage of the intended behaviors. See the project overview and research blog for the motivation and evaluation method.

The rollouts here provide the target-model record on the original tasks. They let users start an AutoDataBench run without regenerating its 144 preparation attempts. They are not the author agents' task-generation traces or the target model's attempts on newly generated tasks.

Dataset at a glance

Suite	Domain	Original tasks	Attempts per task	Total attempts	Solved attempts
AutomationBench	Business workflows through application tools	8	6	48	24
Terminal-Bench	Terminal work and software engineering	8	6	48	5
TB-Science	Scientific computing	8	6	48	7
Total		24	6	144	36

Counts are computed from the packaged rollout/summary.json files. All 144 attempts have a verifier verdict; all 24 tasks use a pass threshold of 1.0. The overall solve rate is 25.0%. These are results on the selected original tasks, not full-suite benchmark scores or AutoDataBench task-authoring scores.

Every transcript identifies the solver as deepseek-v4-pro, running through claude-code 2.1.263. The rubric metadata uses the target-model identifier anthropic-deepseek-v4-pro. Attempts are independent, with no memory carried between them.

Files and format

The release is a gzip-compressed tar archive, rollouts-deepseek-v4-pro.tar.gz (10,996,857 bytes), containing Markdown transcripts and JSON summaries and rubrics.

rollouts-deepseek-v4-pro/
├── README.md
├── automationbench/
├── terminal-bench/
└── tb-science/
    └── <task-name>/
        ├── modes.json
        └── rollout/
            ├── README.md
            ├── summary.json
            ├── attempt-0.md
            ├── attempt-1.md
            ├── attempt-2.md
            ├── attempt-3.md
            ├── attempt-4.md
            └── attempt-5.md

All three suite directories follow the same task layout.

File	Contents
attempt-N.md	Reward, solved status, agent/model metadata and token counts, followed by verifier output and the ordered transcript: user input, recorded reasoning, tool calls and tool results.
summary.json	Task name, attempt count, missing-verdict count, pass threshold, per-attempt rewards and solved arrays, n_solved, and pass_rate.
rollout/README.md	Per-attempt outcome table and guidance for reading the transcripts.
modes.json	Analyst-derived behavioral rubric: model/task metadata and modes with descriptions, examples, and source evidence pointing to attempts and steps.

There are 125 rubric modes across the 24 tasks. Modes can describe failures, detours, or challenges the model handled successfully; a mode is not necessarily a failed verifier assertion.

These are sanitized reading copies. Credentials are masked, and long individual tool outputs may be shortened in the middle while retaining their beginning and end. The archive does not contain the original raw runtime directories or the analyst's container output. Executable task definitions and the evaluation harness are available in the code repository.

Download and inspect

curl -fL --retry 3 -o rollouts-deepseek-v4-pro.tar.gz \
  https://huggingface.co/datasets/LordNoah/AutoDataBench/resolve/main/rollouts-deepseek-v4-pro.tar.gz

tar -xzf rollouts-deepseek-v4-pro.tar.gz

Archive SHA256:

8ed8eb8c6890049e4e4877cd6f9ac5743b8836b3e307e94129d24148c9a8d314

For example, inspect the six attempts for a task and their outcomes:

import json
from pathlib import Path

root = Path("rollouts-deepseek-v4-pro")
rollout = root / "automationbench" / "cash-flow-forecast" / "rollout"
summary = json.loads((rollout / "summary.json").read_text())

print(f"Solved: {summary['n_solved']}/{summary['attempts']}")
for i, (reward, solved) in enumerate(zip(summary["rewards"], summary["solved"])):
    print(f"attempt-{i}.md: reward={reward}, solved={solved}")

transcript = (rollout / "attempt-0.md").read_text()

Use with AutoDataBench

From an AutoDataBench checkout, download and extract the archive as above, then place transcripts and rubrics in their separate expected locations:

for suite in automationbench terminal-bench tb-science; do
  for task_dir in "rollouts-deepseek-v4-pro/$suite/"*/; do
    task=$(basename "$task_dir")
    prep_dir="prep/deepseek-v4-pro/$suite/$task"
    rubric_dir="rubrics/deepseek-v4-pro/$suite/$task"
    mkdir -p "$prep_dir" "$rubric_dir"
    cp -R "$task_dir/rollout" "$prep_dir/"
    cp "$task_dir/modes.json" "$rubric_dir/"
  done
done

Follow the code repository's quick start to configure the endpoint and run an episode.

The author agent receives the original-task transcripts; modes.json is a hidden evaluation rubric and must remain outside the author agent's inputs. To use a different target model, generate its own rollouts and rubrics with run_rollout.py and run_analyst.py rather than reusing this model's behavioral evidence.

Interpretation and use

This release supports reproducing AutoDataBench's preparation stage, inspecting tool-use behavior, and auditing the evidence behind its rubrics. Passing a task's verifier does not mean every decision in its transcript was correct; rubric evidence can capture behavior outside the verifier's checks.

The data covers a selected set of original benchmark tasks and is not a representative sample of all tasks in the three source suites. Keep these original-task records separate from training data when evaluating on the same tasks; some task instructions include explicit benchmark-contamination notices. For source-task attribution and applicable terms, consult the code repository's attribution file.