Add paper links, Figure 1 and citation

a29f310 verified about 12 hours ago

5.54 kB

---
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- arloop
- recurrent
- custom_code
- pretrained
---
# ARLoop-1.2B

📚 [Paper](https://arxiv.org/abs/2611.05234) · 💻 [Code](https://github.com/OSU-NLP-Group/ARLoop)

ARLoop is a fully-open pretrained looped transformer with attention residuals. This release
contains the checkpoints of the 1.2B, 500B-token training run.
It is a **base pretrained model**, intended for text completion, which can be used for mid-training and post-training.

![Figure 1: ARLoop benchmark results](assets/figure1.png)

## Load and generate

```bash
pip install 'transformers==5.12.1' safetensors accelerate
# Install the PyTorch build appropriate for your CUDA environment separately.
```

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "osunlp/ARLoop-1.2B"  # Or the local ARLOOP_hf directory.
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, dtype=torch.bfloat16,
).to("cuda").eval()
inputs = tokenizer("The capital of France is", return_tensors="pt").to("cuda")
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=64, do_sample=False)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

Weights load in BF16 for inference. The stored checkpoint retains the original FP32 weights; routing and normalization statistics use FP32 internally.

## Architecture

| Property | Value |
| --- | --- |
| Physical transformer layers | 24, shared across 3 recurrent passes |
| Effective transformer depth | 72 |
| Hidden / MLP intermediate width | 2048 / 5632 |
| Attention / KV heads | 16 / 4 |
| Context length | 4096 |
| RoPE base | 50000 |
| Attention-residual block size | 24 sublayers |
| Recent residual blocks | 3, with input embedding retained |
| Value embeddings | Shared across alternating physical layers |
| Tokenizer vocabulary | 49152, SmolLM2 tokenizer used in training |
| Output head | Tied to token embeddings |

## License

Model release: [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).

## Training checkpoints

`main` contains the final checkpoint. Each revision below is independently loadable
with `revision="step-010000"` (use the same revision for the model and tokenizer), 
where 10000 denotes the number of optimization steps.

| Optimizer step | Training tokens | Revision |
| ---: | ---: | --- |
| 10,000 | 20,971,520,000 | [step-010000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-010000) |
| 20,000 | 41,943,040,000 | [step-020000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-020000) |
| 30,000 | 62,914,560,000 | [step-030000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-030000) |
| 40,000 | 83,886,080,000 | [step-040000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-040000) |
| 50,000 | 104,857,600,000 | [step-050000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-050000) |
| 60,000 | 125,829,120,000 | [step-060000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-060000) |
| 70,000 | 146,800,640,000 | [step-070000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-070000) |
| 80,000 | 167,772,160,000 | [step-080000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-080000) |
| 90,000 | 188,743,680,000 | [step-090000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-090000) |
| 100,000 | 209,715,200,000 | [step-100000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-100000) |
| 110,000 | 230,686,720,000 | [step-110000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-110000) |
| 120,000 | 251,658,240,000 | [step-120000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-120000) |
| 130,000 | 272,629,760,000 | [step-130000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-130000) |
| 140,000 | 293,601,280,000 | [step-140000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-140000) |
| 150,000 | 314,572,800,000 | [step-150000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-150000) |
| 160,000 | 335,544,320,000 | [step-160000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-160000) |
| 170,000 | 356,515,840,000 | [step-170000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-170000) |
| 180,000 | 377,487,360,000 | [step-180000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-180000) |
| 190,000 | 398,458,880,000 | [step-190000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-190000) |
| 200,000 | 419,430,400,000 | [step-200000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-200000) |
| 210,000 | 440,401,920,000 | [step-210000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-210000) |
| 220,000 | 461,373,440,000 | [step-220000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-220000) |
| 230,000 | 482,344,960,000 | [step-230000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-230000) |
| 238,417 (final) | 499,996,688,384 | [main](https://huggingface.co/osunlp/ARLoop-1.2B/tree/main) / [step-238417](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-238417) |

All 23 scheduled intermediate checkpoints from 10k through 230k are available.

## Citation

```bibtex
@misc{yao2026arloop,
  title={{ARLoop}: Fully-Open Pre-trained Looped Transformers with Attention Residuals},
  author={Yuekun Yao and Yupei Du and Dawei Zhu and Huan Sun},
  year={2026},
  eprint={2611.05234},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2611.05234}
}
```

	---
	language:
	- en
	license: apache-2.0
	library_name: transformers
	pipeline_tag: text-generation
	tags:
	- arloop
	- recurrent
	- custom_code
	- pretrained
	---

	# ARLoop-1.2B

	📚 [Paper](https://arxiv.org/abs/2611.05234) · 💻 [Code](https://github.com/OSU-NLP-Group/ARLoop)

	ARLoop is a fully-open pretrained looped transformer with attention residuals. This release
	contains the checkpoints of the 1.2B, 500B-token training run.
	It is a base pretrained model, intended for text completion, which can be used for mid-training and post-training.

	![Figure 1: ARLoop benchmark results](assets/figure1.png)

	## Load and generate

	```bash
	pip install 'transformers==5.12.1' safetensors accelerate
	# Install the PyTorch build appropriate for your CUDA environment separately.
	```

	```python
	import torch
	from transformers import AutoModelForCausalLM, AutoTokenizer

	model_id = "osunlp/ARLoop-1.2B" # Or the local ARLOOP_hf directory.
	tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
	model = AutoModelForCausalLM.from_pretrained(
	model_id, trust_remote_code=True, dtype=torch.bfloat16,
	).to("cuda").eval()
	inputs = tokenizer("The capital of France is", return_tensors="pt").to("cuda")
	with torch.inference_mode():
	output = model.generate(**inputs, max_new_tokens=64, do_sample=False)
	print(tokenizer.decode(output[0], skip_special_tokens=True))
	```

	Weights load in BF16 for inference. The stored checkpoint retains the original FP32 weights; routing and normalization statistics use FP32 internally.

	## Architecture

	\| Property \| Value \|
	\| --- \| --- \|
	\| Physical transformer layers \| 24, shared across 3 recurrent passes \|
	\| Effective transformer depth \| 72 \|
	\| Hidden / MLP intermediate width \| 2048 / 5632 \|
	\| Attention / KV heads \| 16 / 4 \|
	\| Context length \| 4096 \|
	\| RoPE base \| 50000 \|
	\| Attention-residual block size \| 24 sublayers \|
	\| Recent residual blocks \| 3, with input embedding retained \|
	\| Value embeddings \| Shared across alternating physical layers \|
	\| Tokenizer vocabulary \| 49152, SmolLM2 tokenizer used in training \|
	\| Output head \| Tied to token embeddings \|

	## License

	Model release: [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).

	## Training checkpoints

	`main` contains the final checkpoint. Each revision below is independently loadable
	with `revision="step-010000"` (use the same revision for the model and tokenizer),
	where 10000 denotes the number of optimization steps.

	\| Optimizer step \| Training tokens \| Revision \|
	\| ---: \| ---: \| --- \|
	\| 10,000 \| 20,971,520,000 \| [step-010000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-010000) \|
	\| 20,000 \| 41,943,040,000 \| [step-020000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-020000) \|
	\| 30,000 \| 62,914,560,000 \| [step-030000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-030000) \|
	\| 40,000 \| 83,886,080,000 \| [step-040000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-040000) \|
	\| 50,000 \| 104,857,600,000 \| [step-050000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-050000) \|
	\| 60,000 \| 125,829,120,000 \| [step-060000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-060000) \|
	\| 70,000 \| 146,800,640,000 \| [step-070000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-070000) \|
	\| 80,000 \| 167,772,160,000 \| [step-080000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-080000) \|
	\| 90,000 \| 188,743,680,000 \| [step-090000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-090000) \|
	\| 100,000 \| 209,715,200,000 \| [step-100000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-100000) \|
	\| 110,000 \| 230,686,720,000 \| [step-110000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-110000) \|
	\| 120,000 \| 251,658,240,000 \| [step-120000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-120000) \|
	\| 130,000 \| 272,629,760,000 \| [step-130000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-130000) \|
	\| 140,000 \| 293,601,280,000 \| [step-140000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-140000) \|
	\| 150,000 \| 314,572,800,000 \| [step-150000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-150000) \|
	\| 160,000 \| 335,544,320,000 \| [step-160000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-160000) \|
	\| 170,000 \| 356,515,840,000 \| [step-170000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-170000) \|
	\| 180,000 \| 377,487,360,000 \| [step-180000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-180000) \|
	\| 190,000 \| 398,458,880,000 \| [step-190000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-190000) \|
	\| 200,000 \| 419,430,400,000 \| [step-200000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-200000) \|
	\| 210,000 \| 440,401,920,000 \| [step-210000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-210000) \|
	\| 220,000 \| 461,373,440,000 \| [step-220000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-220000) \|
	\| 230,000 \| 482,344,960,000 \| [step-230000](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-230000) \|
	\| 238,417 (final) \| 499,996,688,384 \| [main](https://huggingface.co/osunlp/ARLoop-1.2B/tree/main) / [step-238417](https://huggingface.co/osunlp/ARLoop-1.2B/tree/step-238417) \|

	All 23 scheduled intermediate checkpoints from 10k through 230k are available.

	## Citation

	```bibtex
	@misc{yao2026arloop,
	title={{ARLoop}: Fully-Open Pre-trained Looped Transformers with Attention Residuals},
	author={Yuekun Yao and Yupei Du and Dawei Zhu and Huan Sun},
	year={2026},
	eprint={2611.05234},
	archivePrefix={arXiv},
	primaryClass={cs.CL},
	url={https://arxiv.org/abs/2611.05234}
	}
	```