Scaling Data Environments from Scratch

Specify the world behind the files. Generate the data. Ground the supervision.

DataWorld-35B-A3B

A reinforcement-learning fine-tune of Qwen/Qwen3.6-35B-A3B (35B total / 3B activated, Apache-2.0), trained with multi-turn agentic RL on the DataWorld task suite through a mini-swe-agent harness. This repository holds the resulting checkpoint, exported to Hugging Face safetensors format.

Format

A standard transformers checkpoint — weights only, no optimizer state and no training configuration. The layout is the base model's:

architecture	Qwen3_5MoeForConditionalGeneration (model_type: qwen3_5_moe), bfloat16
parameters	35B total / 3B activated — 40 layers, 256 experts, top-8 routed
size on disk	70.21 GB across 14 safetensors shards (31,666 tensors)
hidden size	2048 (text)
vocabulary	248,320 tokens
context	262,144 tokens (max_position_embeddings)
attention	hybrid: 3 × linear_attention + 1 × full_attention per block
base model	Qwen/Qwen3.6-35B-A3B

Serving

The architecture, config and weight layout are the base model's, so the official Qwen3.6 deployment recipe applies unchanged — only the model path changes.

python -m sglang.launch_server \
  --model-path Zhanli-Li/DataWorld-35B-A3B \
  --port 8000 --tp-size 8 \
  --mem-fraction-static 0.8 --context-length 262144 \
  --reasoning-parser qwen3 --tool-call-parser qwen3_coder

• Requires sglang >= 0.5.10, vllm >= 0.19.0, or a transformers that knows qwen3_5_moe. The qwen3 reasoning parser and qwen3_coder tool-call parser are what the official card prescribes, and this checkpoint was trained on tool-calling agent trajectories, so keep them enabled.
• The vision tower is included, the MTP block is not. The model.visual.* tensors ship in their own shard, so multimodal serving works as on the base model; mtp.* is absent, so the base card's speculative-decoding recipes do not apply here.
• The long-context (YaRN) overrides apply as they are: SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 / VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 with --context-length / --max-model-len raised to 1,010,000.

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Zhanli-Li/DataWorld-35B-A3B"   # or a local path
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(repo)

Links

• Project page — https://zhanli-li.github.io/Data-World/
• Paper — https://arxiv.org/abs/2610.21656
• Code, data generation pipeline and evaluation harness — https://github.com/Zhanli-Li/Data-World
• Benchmark and training data — https://huggingface.co/collections/Zhanli-Li/dataworld

License

Both the base model and this fine-tune are released under the Apache License 2.0 — see LICENSE and the base model card.