Preview model card for DMax-Coder-MMD

README.md

@@ -0,0 +1,93 @@
+---
+license: apache-2.0
+library_name: transformers
+base_model: Zigeng/DMax-Coder-16B
+base_model_relation: finetune
+datasets:
+  - Zigeng/DMax-LLaDA-2.0-Mini-Code-Trajectories
+tags:
+  - diffusion
+  - dllm
+  - mmd
+  - code
+  - custom_code
+---
+
+# DMax-Coder-MMD
+
+[![License: Apache-2.0](https://img.shields.io/badge/License-Apache--2.0-blue.svg)](https://www.apache.org/licenses/LICENSE-2.0) ![Paper: coming soon](https://img.shields.io/badge/Paper-coming%20soon-b31b1b.svg) [![GitHub: Code](https://img.shields.io/badge/GitHub-Code-181717.svg?logo=github)](https://github.com/yandex-research/dlm-mmd)
+
+**Representation-Space MMD for Diffusion Language Models**
+
+DMax-Coder-MMD is a 16B diffusion language model for code generation,
+obtained by MMD post-training of
+[DMax-Coder-16B](https://huggingface.co/Zigeng/DMax-Coder-16B).
+It builds on LLaDA2.0-mini and uses DMax's hybrid masked–uniform block diffusion.
+
+The post-training objective minimizes Maximum Mean Discrepancy (MMD) between
+model samples and reference responses, measured in the representation space of
+a frozen diffusion language model. The project reports increased tokens per
+forward pass while maintaining or improving accuracy on the benchmarks below.
+
+## Reference results
+
+Results reported in the project README, with decoding threshold **0.9**.
+Each entry is **accuracy (%) / tokens per forward pass (TPF)**.
+Baseline results are attributed to the original DMax paper in the project README.
+
+| Method | HumanEval-Instruct | MBPP-Instruct |
+| --- | :---: | :---: |
+| DMax-Coder | 83.5 / 7.36 | 79.2 / 5.86 |
+| **DMax-Coder-MMD** | **85.9** / **8.07** | **83.0** / **6.10** |
+
+Results can vary with hardware, tensor parallelism, and library versions.
+TPF measures decoding parallelism; wall-clock speed also depends on the runtime.
+
+## Evaluation and inference
+
+Use DMax's dInfer evaluation pipeline. After following the evaluation environment
+setup in the [project README](https://github.com/yandex-research/dlm-mmd#installation),
+run from the project repository root:
+
+```bash
+conda activate dinfer
+DOMAIN=code MODEL_PATH=yresearch/DMax-Coder-MMD bash scripts/eval.sh
+```
+
+The project's code evaluation uses threshold **0.9** and evaluates
+HumanEval-Instruct and MBPP-Instruct.
+
+The checkpoint includes nine Safetensors shards with BF16 parameters and FP32
+router bias buffers, plus its tokenizer, chat template, and custom model code.
+Direct Transformers loading requires `trust_remote_code=True`. Pass
+`dtype=torch.bfloat16` for BF16 loading; the current configuration declares FP32.
+
+Generation requires the DMax/dInfer diffusion decoder. The included custom
+backbone does not implement the standard Transformers `.generate()` interface.
+
+## License and acknowledgements
+
+The model follows the **Apache-2.0** license of the
+[DMax-Coder base checkpoint](https://huggingface.co/Zigeng/DMax-Coder-16B)
+and [LLaDA2.0-mini](https://huggingface.co/inclusionAI/LLaDA2.0-mini).
+The included model implementation retains its Apache-2.0 notices.
+The separate MMD training repository is MIT-licensed, with Apache-2.0 third-party components.
+
+We thank the authors of [DMax](https://github.com/czg1225/DMax),
+[LLaDA2.0-mini](https://huggingface.co/inclusionAI/LLaDA2.0-mini), and
+[dInfer](https://github.com/inclusionAI/dInfer) for releasing their models, data, and code.
+
+## Citation
+
+If you find this work useful in your research, please consider citing:
+
+```bibtex
+@article{drobyshevskiy2026mmd,
+  title  = {Representation-Space MMD for Diffusion Language Models},
+  author = {Drobyshevskiy, Ilya and Sudakov, Ilia and Semenov, Maksim and Kuznedelev, Denis and
+            Ignatov, Maksim and Temirchev, Pavel and Balagansky, Nikita and
+            Meshchaninov, Viacheslav and Gushchin, Nikita and Baranchuk, Dmitry},
+  journal = {arXiv preprint TODO},
+  year    = {2026}
+}
+```