Add pipeline tag and link to paper/code
#1
by nielsr HF Staff - opened
README.md
CHANGED
|
@@ -1,4 +1,9 @@
|
|
| 1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
language:
|
| 3 |
- en
|
| 4 |
- zh
|
|
@@ -13,23 +18,39 @@ language:
|
|
| 13 |
- vi
|
| 14 |
- ar
|
| 15 |
license: apache-2.0
|
|
|
|
| 16 |
tags:
|
| 17 |
- dllm
|
| 18 |
- diffusion
|
| 19 |
- rwkv
|
| 20 |
- llm
|
| 21 |
-
- text_generation
|
| 22 |
-
datasets:
|
| 23 |
-
- allenai/tulu-3-sft-mixture
|
| 24 |
-
- Jackrong/GLM-5.1-Reasoning-1M-Cleaned
|
| 25 |
-
- angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k
|
| 26 |
-
base_model: BlinkDL/rwkv7-g1
|
| 27 |
---
|
| 28 |
|
| 29 |
# Triplet-Block Diffusion RWKV
|
| 30 |
|
| 31 |
-
This repository contains the checkpoint of B3D-RWKV, a
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
|
| 33 |
-
|
| 34 |
|
| 35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
base_model: BlinkDL/rwkv7-g1
|
| 3 |
+
datasets:
|
| 4 |
+
- allenai/tulu-3-sft-mixture
|
| 5 |
+
- Jackrong/GLM-5.1-Reasoning-1M-Cleaned
|
| 6 |
+
- angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k
|
| 7 |
language:
|
| 8 |
- en
|
| 9 |
- zh
|
|
|
|
| 18 |
- vi
|
| 19 |
- ar
|
| 20 |
license: apache-2.0
|
| 21 |
+
pipeline_tag: text-generation
|
| 22 |
tags:
|
| 23 |
- dllm
|
| 24 |
- diffusion
|
| 25 |
- rwkv
|
| 26 |
- llm
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
---
|
| 28 |
|
| 29 |
# Triplet-Block Diffusion RWKV
|
| 30 |
|
| 31 |
+
This repository contains the checkpoint of **B3D-RWKV**, a 7.2B-parameter RWKV language model presented in the paper [Triplet-Block Diffusion RWKV](https://arxiv.org/abs/2605.25969).
|
| 32 |
+
|
| 33 |
+
B3D-RWKV is a diffusion RWKV variant that integrates the model's $O(L)$ inference efficiency with parallel, bidirectional discrete-diffusion through a *triplet-block layout* method. It reaches comparable accuracy on an 8-task suite versus existing models while significantly outperforming baselines in decoding throughput with an average of **1.6×** speedup.
|
| 34 |
+
|
| 35 |
+
- **Paper:** [Triplet-Block Diffusion RWKV](https://arxiv.org/abs/2605.25969)
|
| 36 |
+
- **Code:** [GitHub Repository](https://github.com/leonardodalinky/B3D-RWKV)
|
| 37 |
+
|
| 38 |
+
## Usage
|
| 39 |
+
|
| 40 |
+
For usage, please see the B3D-RWKV [infer](https://github.com/leonardodalinky/B3D-RWKV/tree/main/infer) and [serve](https://github.com/leonardodalinky/B3D-RWKV/tree/main/infer/serve) directories in the official repository for instructions on how to run inference and serve the model.
|
| 41 |
+
|
| 42 |
+
Note: This checkpoint is a supervised fine-tuned (SFT) version of `rwkv7-g1f-7.2B`.
|
| 43 |
|
| 44 |
+
## Citation
|
| 45 |
|
| 46 |
+
```bibtex
|
| 47 |
+
@misc{lin2026tripletblockdiffusionrwkv,
|
| 48 |
+
title={Triplet-Block Diffusion RWKV},
|
| 49 |
+
author={Ke Lin and Yiyang Luo and Zhaolong Su and Yunya Song and Anyi Rao},
|
| 50 |
+
year={2026},
|
| 51 |
+
eprint={2605.25969},
|
| 52 |
+
archivePrefix={arXiv},
|
| 53 |
+
primaryClass={cs.CL},
|
| 54 |
+
url={https://arxiv.org/abs/2605.25969},
|
| 55 |
+
}
|
| 56 |
+
```
|