Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- language-model
|
| 7 |
+
- flow-matching
|
| 8 |
+
- diffusion
|
| 9 |
+
- hypersphere
|
| 10 |
+
- discrete-diffusion
|
| 11 |
+
datasets:
|
| 12 |
+
- tinygsm
|
| 13 |
+
- openwebtext
|
| 14 |
+
library_name: pytorch
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Language Modeling with Hyperspherical Flows
|
| 18 |
+
|
| 19 |
+
By [Justin Deschenaux](https://jdeschena.com) and [Caglar Gulcehre](https://www.caglar.ai).
|
| 20 |
+
|
| 21 |
+
[](https://arxiv.org/abs/2605.11125)
|
| 22 |
+
[](https://jdeschena.com/blog/sfm)
|
| 23 |
+
[](https://github.com/jdeschena/s-flm)
|
| 24 |
+
|
| 25 |
+
This repo hosts the pretrained checkpoints for **Language Modeling with Hyperspherical Flows** (𝕊-FLM). For the abstract, training/sampling code, and reproduction scripts, see the companion code repo: [`jdeschena/s-flm`](https://github.com/jdeschena/s-flm).
|
| 26 |
+
|
| 27 |
+
# Checkpoints
|
| 28 |
+
|
| 29 |
+
𝕊-FLM and the baselines we compare against (AR, MDLM, Duo, FLM, CANDI), trained on **TinyGSM** (250k steps, SmolLM-135M tokenizer) and **OpenWebText** (1M steps, GPT-2 tokenizer).
|
| 30 |
+
|
| 31 |
+
```
|
| 32 |
+
tinygsm/{ar,mdlm,duo}.ckpt
|
| 33 |
+
tinygsm/candi/{lr3e-4,lr1e-3}.ckpt
|
| 34 |
+
tinygsm/flm/{default,caps}.ckpt
|
| 35 |
+
tinygsm/sfm/{sphere_dit_truncated_fixed_no_renorm,
|
| 36 |
+
sphere_dit_truncated_adaptive_no_renorm,
|
| 37 |
+
sphere_arch_truncated_adaptive_no_renorm}.ckpt
|
| 38 |
+
|
| 39 |
+
owt/{ar,mdlm,duo,flm,sfm}.ckpt
|
| 40 |
+
```
|
| 41 |
+
|
| 42 |
+
```bash
|
| 43 |
+
huggingface-cli download jdeschena/s-flm tinygsm/duo.ckpt --local-dir ./checkpoints
|
| 44 |
+
```
|
| 45 |
+
|
| 46 |
+
Loading and sampling are handled by the code repo — see [`jdeschena/s-flm`](https://github.com/jdeschena/s-flm) for the scripts.
|
| 47 |
+
|
| 48 |
+
# Citation
|
| 49 |
+
|
| 50 |
+
```
|
| 51 |
+
@misc{deschenaux2026languagemodelinghypersphericalflows,
|
| 52 |
+
title={Language Modeling with Hyperspherical Flows},
|
| 53 |
+
author={Justin Deschenaux and Caglar Gulcehre},
|
| 54 |
+
year={2026},
|
| 55 |
+
eprint={2605.11125},
|
| 56 |
+
archivePrefix={arXiv},
|
| 57 |
+
primaryClass={cs.LG},
|
| 58 |
+
url={https://arxiv.org/abs/2605.11125},
|
| 59 |
+
}
|
| 60 |
+
```
|