|
--- |
|
language: |
|
- de |
|
- en |
|
license: llama2 |
|
library_name: transformers |
|
tags: |
|
- llama2 |
|
- deutsch |
|
- german |
|
- seedbox |
|
datasets: |
|
- seedboxai/multitask_german_examples_32k |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
|
|
 |
|
|
|
|
|
|
|
# KafkaLM-7B-DARE_TIES-LaserRMT-QLoRA-DPO-v0.5 |
|
|
|
**KafkaLM 7b** is a Mistral 7b model - further pre-trained on a large German dataset from Björn Plüster and LAION. [leo-mistral-hessianai-7b](https://huggingface.co/LeoLM/leo-mistral-hessianai-7b) - which was finetuned on an ensemble of popular high-quality open-source instruction sets (translated from English to German). |
|
|
|
KafkaLM 7b is a [Seedbox](https://huggingface.co/seedboxai) project trained by [Dennis Dickmann](https://huggingface.co/doubledsbv). |
|
|
|
**Why Kafka?** |
|
The models are proficient, yet creative, and have some tendencies to linguistically push boundaries 😊 |
|
|
|
|
|
|
|
## THE MODEL CAN BE TESTET HERE [Kafka-7B HF Space](https://huggingface.co/spaces/doubledsbv/Kafka-7B-DARE-TIES-QLoRa-LaserRMT-DPO) |
|
|
|
|
|
|
|
## Model Details |
|
|
|
The purpose of releasing the **KafkaLM series** is to contribute to the German AI community with a set of fine-tuned LLMs that are easy to use in everyday applications across a variety of tasks. |
|
|
|
The main goal was to provide LLMs proficient in German, especially to be used in German-speaking business contexts where English alone is not sufficient. |
|
|
|
|
|
## DPO Training with laserRMT w/ Q-Lora |
|
|
|
Based on the brilliant work from [laserRMT](https://github.com/cognitivecomputations/laserRMT/) team, I used the SNR implementation for identifying candiate layers to be used for the DPO training. |
|
|
|
### Dataset |
|
|
|
I used a 8k filtered version of the following [seedboxai/multitask_german_examples_32k](https://huggingface.co/datasets/seedboxai/multitask_german_examples_32k) |
|
|
|
### Prompt Format |
|
|
|
|
|
This model follows the subsequent prompt format: |
|
|
|
``` |
|
<|system|> |
|
Du bist ein freundlicher und hilfsbereiter KI-Assistent. Du beantwortest Fragen faktenorientiert und präzise, ohne dabei relevante Fakten auszulassen.</s> |
|
<|user|> |
|
Welche Möglichkeiten der energetischen Sanierung habe ich neben Solar und Energiespeicher?</s> |
|
<|assistant|> |
|
``` |
|
|
|
|
|
|
|
|
|
## 🧩 Configuration |
|
|
|
```yaml |
|
models: |
|
- model: mistralai/Mistral-7B-v0.1 |
|
# no parameters necessary for base model |
|
- model: seedboxai/KafkaLM-7B-German-V0.1 |
|
parameters: |
|
density: 0.65 |
|
weight: 0.50 |
|
- model: mlabonne/Monarch-7B |
|
parameters: |
|
density: 0.60 |
|
weight: 0.30 |
|
- model: mayflowergmbh/Wiedervereinigung-7b-dpo-laser |
|
parameters: |
|
density: 0.60 |
|
weight: 0.20 |
|
merge_method: dare_ties |
|
base_model: mistralai/Mistral-7B-v0.1 |
|
parameters: |
|
int8_mask: true |
|
dtype: bfloat16 |
|
random_seed: 0 |
|
``` |
|
|
|
## 💻 Usage |
|
|
|
```python |
|
!pip install -qU transformers accelerate |
|
|
|
from transformers import AutoTokenizer |
|
import transformers |
|
import torch |
|
|
|
model = "seedboxai/KafkaLM-7B-DARE_TIES-LaserRMT-QLoRA-DPO-v0.5" |
|
messages = [{"role": "user", "content": "Was ist der Sinn des Lebens?"}] |
|
|
|
tokenizer = AutoTokenizer.from_pretrained(model) |
|
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) |
|
pipeline = transformers.pipeline( |
|
"text-generation", |
|
model=model, |
|
torch_dtype=torch.float16, |
|
device_map="auto", |
|
) |
|
|
|
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95) |
|
print(outputs[0]["generated_text"]) |
|
``` |
|
|
|
## Disclaimer |
|
|
|
The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. |
|
This model should only be used for research purposes. The original Llama2 license and all restrictions of datasets used to train this model apply. |
