---
license: mit
library_name: pytorch
pipeline_tag: text-generation
tags:
  - reinforcement-learning
  - flow-matching
  - continuous-language-model
  - reasoning
---

# CFRL checkpoints: RL post-training of the ELF-B continuous flow language model

Checkpoints for the paper **CFRL** (reinforcement learning for continuous flow language models). Every file is a
PyTorch weights file of ELF-B (105M parameters), the embedded-language-flow model of
[ELF](https://huggingface.co/embedded-language-flows), fine-tuned on four program-verified reasoning tasks
(Countdown, Game24, Sudoku, Knights-and-Knaves) and then post-trained with one of three RL variants.

The code that trains, evaluates and loads these files is the `cfrl-release` repository (`cfrl/common.py:load_weights`,
`scripts/evaluate.sh`). This repository holds only the weights.

## Layout

| folder | files | what |
|---|---|---|
| `base/` | `<task>.pt` | task-SFT ELF-B, the starting point of every RL run |
| `critic/` | `<task>_1d.pt`, `<task>_2d.pt` | critics (denoiser copy + MLP head) used by CFRL-SVG-1D / CFRL-SVG-2D |
| `grr/` | `<task>.pt` | CFRL-GRR policy, one per task |
| `svg1d/` | `<task>.pt` | CFRL-SVG-1D policy, one per task |
| `svg2d/` | `<task>.pt` | CFRL-SVG-2D policy, one per task |

`<task>` is one of `countdown`, `game24`, `sudoku`, `kk`. Model files are dicts with `params`, `ema_params1` and `step`
(EMA weights are used when present); critic files hold `trunk`, `head` and `mode`.

## Results

Test accuracy, Table 1. 

| task | base | CFRL-GRR | CFRL-SVG-1D | CFRL-SVG-2D |
|---|---|---|---|---|
| Countdown | 42.8 | 42.6 | 47.4 | 49.7 |
| Sudoku | 64.0 | 67.9 | 64.1 | 68.7 |
| Knights-and-Knaves | 68.2 | 71.4 | 76.5 | 77.5 |
| Game24 | 44.0 | 47.4 | 51.2 | 58.2 |


## Usage

```bash
# in the cfrl-release repository, with its environment
hf download tufts-ml/CFRL --local-dir checkpoints
scripts/evaluate.sh countdown checkpoints/svg2d/countdown.pt
scripts/train_svg.sh countdown 2d        # CFRL-SVG-2D from base/countdown.pt with critic/countdown_2d.pt
```

The files are plain `torch.save` dicts and load with `torch.load(path, map_location="cpu", weights_only=False)`.

## License

MIT, as the code release. ELF-B itself is released under the MIT license of its authors.

## Citation

```bibtex
@article{cfrl2026,
  title   = {CFRL: Reinforcement Learning for Continuous Flow Language Models},
  year    = {2026}
}
```