---
datasets:
  - togethercomputer/RedPajama-Data-V2
language:
  - de
pipeline_tag: text-generation
library_name: transformers
license: other

---

# LLäMmlein 120M


LLäMmlein 120M is a German LLaMa model trained from scratch using our adapted [Tinyllama](https://github.com/jzhang38/TinyLlama) codebase on the German portion of [RedPajama V2](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-V2).
To enhance data quality, we additionally deduplicated the dataset on paragraph level and filtered it using a token-to-word ratio filter. The resulting dataset can be found [here](https://huggingface.co/datasets/LSX-UniWue/LLaMmlein-Dataset). 

We provide three model sizes:

* [LLäMmlein 7B](https://huggingface.co/LSX-UniWue/LLaMmlein_7B) 

* [LLäMmlein 1B](https://huggingface.co/LSX-UniWue/LLaMmlein_1B) 

* [LLäMmlein 120M](https://huggingface.co/LSX-UniWue/LLaMmlein_120M) ← You are here  


Find more details on our page our [page](https://www.informatik.uni-wuerzburg.de/datascience/projects/nlp/llammlein/) and our [preprint](https://arxiv.org/abs/2411.11171)!


### Usage

You can use LLäMmlein with the `transformers` library. 
(Optional: install `flash-attn` to achieve highest efficiency.)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "LSX-UniWue/LLaMmlein_120M"
tokenizer = AutoTokenizer.from_pretrained("LSX-UniWue/LLaMmlein_120M")
model = AutoModelForCausalLM.from_pretrained("LSX-UniWue/LLaMmlein_120M")
```


### Intermediate Checkpoints
In addition to the final model checkpoint, we publish intermediate checkpoints throughout the full training process as unique branches in this repository. 
A specific checkpoint can be loaded like this: 

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "LSX-UniWue/LLaMmlein_120M"
revision = "iter-00420000-ckpt"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
model = AutoModelForCausalLM.from_pretrained(model_id, revision=revision)
```

Next to the model itself each branch contains all datapoints that were used to train the model up to that point.
In the correspinding folder, named after the checkpoint, you can find several `.log` files (depending on the number of GPUs) of the following format: 

```json
{"time": 1739809392.679516, 
  "iter_num": 0, 
  "data_id": ["sha1:EDQMBYDCYBLDAZH3MGYM276BM2DEHPPJ", "sha1:SAJCI75DRHZZFGQORV66NB5FVWUAVLFH", "sha1:7RBZV2MCEM4TUGBBWGTFQAKTWUOGETZU", "sha1:234M32IMLZF7455AKOFWDP6HT6YXAYB4", "sha1:2BIZ7LLSHRK5GUGPZM2GM55APTDKBUG2", "sha1:OF7OI77ZT7ROXGMB6LL4RSRANX7REAYK", "sha1:LGPUOCOV3MKETI5F3IHVGZPD4M26NNJL", "sha1:SHIHUW7FJTP5YHFFV2JZ2CAHUVMKK7XG"], 
  "file_id": [0, 0, 0, 0, 0, 0, 0, 0], 
  "process_rank": 0}
```


Note: Our earlier models from the paper, which do not include data logging, are available at:
* [LLäMmlein 1B prerelease](https://huggingface.co/LSX-UniWue/LLaMmlein_1B_prerelease) 

* [LLäMmlein 120M prerelease](https://huggingface.co/LSX-UniWue/LLaMmlein_120M_prerelease) 



### License
We release the LLäMmlein models under a research-only RAIL-M license. See [license.md](./license.md) for details.