---
license: other
language:
- en
tags:
- midtraining
- code
- agentic
- software-engineering
- multi-agent
task_categories:
- text-generation
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: text
    dtype: string
  - name: source
    dtype: string
  splits:
  - name: train
    num_bytes: 1610042422
    num_examples: 727337
  download_size: 806246956
  dataset_size: 1610042422
---

# CooperData v3 — Midtraining Blend (Qwen3.5-9B cooperative SWE agents)

All-token **midtraining** mixture that bridges `Qwen/Qwen3.5-9B` (instruct) toward the
cooperative multi-agent SWE-coding SFT distribution. **One document per row** (`text`, tagged by
`source`) — NOT packed — so `trl.SFTTrainer(packing=False)` tokenizes per-doc and the Gated-DeltaNet
recurrence stays per-document. **~390M tokens.**

## Composition
| source | tokens | share |
|---|---|---|
| `web` | 210.0M | 54% |
| `math` | 55.0M | 14% |
| `instruct` | 25.0M | 6% |
| `coop` | 80.0M | 21% |
| `agentic` | 20.0M | 5% |

- **General (~81%)** — anti-forgetting bulk + the distant domains (code/math) midtraining helps most:
  `web`/`math`/`instruct` = [allenai/dolmino-mix-1124](https://huggingface.co/datasets/allenai/dolmino-mix-1124) (`dclm`/`math`/`flan`), OLMo-2's curated midtraining mix.
- **Bridge (~19%, coop-dominant)** — agentic data resembling the SFT target:
  - `coop`: our cooperative multi-agent SWE trajectories (mini-swe-agent, **structured tool-calling**), rendered with the model's **real** chat template → native `<tool_call>`/`<tool_response>` (matches the `qwen3_coder` serving parser). This also carries the "good collaborator" signal — the agents coordinate as colleagues on a shared codebase.
  - `agentic`: [ricdomolm/mini-coder-trajs-400k](https://huggingface.co/datasets/ricdomolm/mini-coder-trajs-400k), verified rollouts (bash-in-content style — midtrain *exposure* only; the SFT stage uses coop tool-calling exclusively).

## Methodology
Grounded in ["Midtraining Bridges Pretraining and Posttraining Distributions" (2510.14865)](https://arxiv.org/abs/2510.14865): because specialized data is introduced *late* (a finished instruct model), the conservative-weight finding (high specialized weight late → catastrophic forgetting) dictates a **general-dominant** mix (~81% general / ~19% bridge). Heavy agentic specialization is left to SFT.

## Loss / usage
All-token (full LM loss), `packing=False`:
```python
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset
ds = load_dataset("CooperBench/cooperdata-v3-midtrain-blend", split="train")
SFTTrainer(model="Qwen/Qwen3.5-9B", train_dataset=ds,
           args=SFTConfig(dataset_text_field="text", packing=False, max_length=8192))
```
Inherits the licenses of its constituent datasets.