---
license: other
license_name: microtensor-dataset-terms
license_link: https://www.microtensor.cloud/terms
pretty_name: Microtensor agent sessions
language:
- en
tags:
- agents
- context-compression
- compaction
- code
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: dev
    path: data/dev-*.parquet
---

# Microtensor agent sessions

Coding agent sessions for the `ctx_compress` arena on Microtensor (Bittensor testnet netuid 576). Each row is one compaction request in the arena's exact format: the compaction instruction as the system message and the session, rendered canonically inside `<session>` tags, as the user message. The label holds the facts and probe questions the grader checks.

Every session is generated by Microtensor. A seed coins a Python project with one injected bug, and a scripted agent works the task: it lists the tree, searches, reads modules and tests, runs pytest, often tries a wrong fix and reverts it, applies the right fix and adds a regression test. Every tool result is the real output of the real command. Nothing here comes from a public dataset.

| | |
|---|---|
| Sessions | 2,000 |
| Rows | 3,981 (each session cut near its end, and a twin cut near its middle: 1,981 twins) |
| Input tokens | 6,381 to 122,704, median 45,439, Granite 4.0 H Micro tokenizer |
| Arena input cap | 122,880 tokens |
| Output | nine heading markdown summary within `budget_tokens` |

Hidden evaluation items come from the same generator under a secret seed and are never published.

## Fields

| Field | Meaning |
|---|---|
| `item_id` | row id; a twin's id ends in `-twin` |
| `twin_of` | the full item a twin was cut from, empty otherwise |
| `family` | the generated project name |
| `system`, `user` | the request exactly as a miner's model receives it |
| `input_tokens`, `budget_tokens` | Granite token count of the request and the output budget it states |
| `cut`, `steps` | the step the session was cut at and its full length |
| `facts` | JSON list of files, errors and symbols the summary must keep |
| `probes` | JSON list of continuation questions with accepted answers |
| `image` | `synth:<seed>@<scale>`; the arenas package rebuilds the buggy project from it |


## Use

```python
from datasets import load_dataset

rows = load_dataset("microtensor-io/agent-sessions", split="dev")
```

There are no reference summaries. Write targets with a teacher model and keep those the arena grader passes; the miner guide shows the loop. Make more sessions with your own seed:

```
pip install "microtensor-arenas[compress] @ git+https://github.com/microtensor-io/microtensor-arenas"
mta dev --arena ctx_compress --count 2000 --seed 7 --out dev/
```

## Licence

Microtensor dataset terms, see https://www.microtensor.cloud/terms.