README.md

---
license: odc-by
pretty_name: ACT-2.5B
language:
  - en
  - zh
size_categories:
  - 10K<n<100K
task_categories:
  - text-generation
tags:
  - agents
  - coding-agent
  - trajectories
  - reasoning
  - real-user-data
configs:
  - config_name: default
    data_files:
      - split: train
        path: parquet/sessions.parquet
---

# ACT-2.5B

**Preview release: 183 sessions.** This repository holds a preview of the dataset
described in a paper currently under double-blind review, so that the format and the
content can be inspected. It was drawn as a 200-session sample, of which 17 were withheld
during pre-release screening: 15 after a review of embedded images and 2 after an
identifier scan. The full subset of 26,999 sessions is released on
acceptance, once the redaction audit is complete. The paper's appendix carries the full
datasheet.

## What the full dataset contains

**ACT-2.5B** is a dataset of **26,999 real coding-agent sessions (2.50B tokens)**,
contributed by developers who authorized their sessions for publication, redacted, and
released under ODC-BY. Each session records a complete interaction: the system prompt, the
tool definitions, and the alternating sequence of user messages, assistant messages, tool
calls, and tool results, spanning six model generations.

What defines the dataset is what it lacks. Across **1.65M assistant turns** it carries only
**15.3 non-empty reasoning blocks per 100 turns**, and **34% of sessions contain no reasoning
at all**. The actions are recorded; the reasons are not. A model trained on such data learns
to imitate what was done rather than work out what to do.

## Collection and consent

Sessions come from developers who accepted a data collection and use authorization before
contributing. The authorization names public release explicitly among its permitted uses,
requires contributors to affirm they are at least 18 (or the local age of majority, if
higher), and commits that no contributor identity is published. Every contributor
represented here accepted it. The operative clauses are reproduced in the paper's appendix.

## Processing

Credentials and contact details are redacted, and home-directory paths are pseudonymized
consistently so that path structure survives. Sessions touching repositories in SWE-bench
Verified are excluded, and near-duplicate sessions are labelled with cluster identifiers
rather than removed. Every thinking block is labelled by whether it is empty and, where
determinable, by cause.

## Code

An anonymized mirror of the experiment and synthesis code accompanies the paper under
review.

## Files

| Path | Contents |
|---|---|
| `data/sessions.jsonl` | The sessions, one JSON object per line. This is the canonical file; `CHECKSUMS.sha256` covers it. |
| `parquet/sessions.parquet` | The same 183 sessions rendered for the dataset viewer. `session_id`, `model`, `thinking_effort`, `n_messages` and `n_tools` are columns; `system`, `tools` and `messages` are serialized as JSON strings. |
| `VERIFICATION.md`, `EXCLUSIONS.md`, `STATS.json` | The pre-release screening record. |
| `consent/` | The authorization and the privacy policy, with identifying names replaced. |

The viewer needs one schema per column, and two fields in the raw file do not have one:
`system` is a list in 181 sessions and a string in 2, and its list elements carry
`cache_control` in some sessions and not in others. Serializing those fields as JSON strings
in the Parquet rendering keeps every byte while giving the viewer a stable schema. Round-trip
was checked on all 183 sessions.

## Loading

```python
from datasets import load_dataset

# the viewer rendering, nested fields as JSON strings
ds = load_dataset("submit2331/ACT-2.5B", split="train")

# the canonical file
import json
rows = [json.loads(l) for l in open("sessions.jsonl", encoding="utf-8")]
```

## Licence

ODC-BY.