Add files using upload-large-folder tool
Browse files
README.md
CHANGED
|
@@ -1,46 +1,43 @@
|
|
| 1 |
---
|
| 2 |
-
license:
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
---
|
| 5 |
|
| 6 |
-
|
| 7 |
-
On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models
|
| 8 |
-
</h1>
|
| 9 |
|
| 10 |
-
|
| 11 |
|
| 12 |
-
|
| 13 |
-
<a href="https://xiangyue9607.github.io">Xiang Yue</a>
|
| 14 |
|
| 15 |
-
|
|
|
|
| 16 |
|
| 17 |
-
|
| 18 |
|
| 19 |
-
|
| 20 |
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
|
|
|
| 24 |
|
| 25 |
-
|
| 26 |
|
|
|
|
|
|
|
| 27 |
|
| 28 |
-
|
|
|
|
| 29 |
|
| 30 |
-
|
|
|
|
|
|
|
| 31 |
|
| 32 |
-
##
|
| 33 |
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
```bibtex
|
| 37 |
-
@misc{zhang2025interplaypretrainingmidtrainingrl,
|
| 38 |
-
title={On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models},
|
| 39 |
-
author={Charlie Zhang and Graham Neubig and Xiang Yue},
|
| 40 |
-
year={2025},
|
| 41 |
-
eprint={2512.07783},
|
| 42 |
-
archivePrefix={arXiv},
|
| 43 |
-
primaryClass={cs.CL},
|
| 44 |
-
url={https://arxiv.org/abs/2512.07783},
|
| 45 |
-
}
|
| 46 |
-
```
|
|
|
|
| 1 |
---
|
| 2 |
+
license: other
|
| 3 |
+
library_name: transformers
|
| 4 |
+
tags:
|
| 5 |
+
- reasoning
|
| 6 |
+
- extrapolation
|
| 7 |
+
- synthetic-data
|
| 8 |
+
- transformers
|
| 9 |
---
|
| 10 |
|
| 11 |
+
# Interplay-LM Extrapolation RL Models
|
|
|
|
|
|
|
| 12 |
|
| 13 |
+
This repository is organized by experiment setting. Each top-level directory corresponds to one pretraining mixture used in the extrapolation experiments.
|
| 14 |
|
| 15 |
+
Within each setting:
|
|
|
|
| 16 |
|
| 17 |
+
- `base/` stores the base model used to initialize RL.
|
| 18 |
+
- `rl/` stores the final RL checkpoints for each experiment variant.
|
| 19 |
|
| 20 |
+
Only inference-relevant Hugging Face files are included.
|
| 21 |
|
| 22 |
+
## Included settings
|
| 23 |
|
| 24 |
+
- `id2-10_0.2easy_0.3medium_0.5hard`
|
| 25 |
+
- `id2-10_0.5easy_0.3medium_0.2hard`
|
| 26 |
+
- `id2-10_0.4995easy_0.4995medium_0.001hard`
|
| 27 |
+
- `id2-10_0.475easy_0.475medium_0.05hard`
|
| 28 |
|
| 29 |
+
## Load
|
| 30 |
|
| 31 |
+
```python
|
| 32 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 33 |
|
| 34 |
+
repo_id = "Interplay-LM-Reasoning/extrapolation_rl"
|
| 35 |
+
subdir = "id2-10_0.5easy_0.3medium_0.2hard/rl/op11-14_uniform"
|
| 36 |
|
| 37 |
+
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder=subdir)
|
| 38 |
+
model = AutoModelForCausalLM.from_pretrained(repo_id, subfolder=subdir)
|
| 39 |
+
```
|
| 40 |
|
| 41 |
+
## Reference
|
| 42 |
|
| 43 |
+
- Zhang, Charlie; Neubig, Graham; Yue, Xiang. "On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models." arXiv:2512.07783 (2025).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|