license: other
language:
- en
tags:
- midtraining
- code
- agentic
- software-engineering
- multi-agent
task_categories:
- text-generation
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 | role |
|---|---|---|---|
web |
210.0M | 54% | general |
math |
55.0M | 14% | general |
instruct |
25.0M | 6% | general |
coop |
60.0M | 15% | bridge |
swe_smith |
15.0M | 4% | bridge |
nebius |
15.0M | 4% | bridge |
social |
10.0M | 3% | bridge |
General ~74% / bridge ~26% (coop-dominant). Agentic/social sources are rendered with
the model's real chat template (so coop tool-calls become native <tool_call>/<tool_response>,
matching the qwen3_coder serving parser).
Methodology
Grounded in "Midtraining Bridges Pretraining and Posttraining Distributions" (2510.14865): specialized data is introduced late (a finished instruct model), so the conservative-weight finding (high specialized weight late -> catastrophic forgetting) dictates a general-dominant mix. The bridge is coop-dominant — near-domain agentic-coding on a coding model (low forgetting risk) and the exact SFT target. Heavy agentic specialization is left to SFT.
Usage (all-token, packing=False)
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))
Build
coopertrain/train/modal/datamix/ (sources + mixture + cleaning). Inherits the licenses of its
constituent datasets (Dolmino, mini-coder-trajs, OdysSim) + our coop trajectories.