---
license: apache-2.0
base_model: mistralai/Mistral-7B-v0.1
---

# Model Card for Mistral7B-v0.1-coco-caption-de

This model is a fine-tuned model of the Mistral7B-v0.1 completion model and meant to produce german COCO like captions.

The [coco-karpathy-opus-de dataset](https://huggingface.co/datasets/Jotschi/coco-karpathy-opus-de) was used to tune the model for german image caption generation.

## Model Details

### Prompt format

The completion model is trained with the prompt prefix `Bildbeschreibung: `

Examples:

```xml
>>> Bildbeschreibung: 
2 Hunde sitzen auf einer Bank neben einer Pflanze

>>> Bildbeschreibung: Wasser
fall und Felsen vor dem Gebäude mit Blick auf den Fluss.

>>> Bildbeschreibung: Ein grünes Auto mit roten 
 Reflektoren parkte auf dem Parkplatz.
```

### Model Description

- **Developed by:** [Jotschi](https://huggingface.co/Jotschi)
- **License:** [Apache License](https://www.apache.org/licenses/LICENSE-2.0)
- **Finetuned from model:** [Mistral7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)

## Uses

The model is meant to be used in conjunction with a [BLIP2](https://huggingface.co/docs/transformers/model_doc/blip-2) Q-Former to enable image captioning, visual question answering (VQA) and chat-like conversations.

## Training Details

The preliminary [training script](https://github.com/Jotschi/lavis-experiments/tree/master/mistral-deepspeed) uses PEFT and DeepSpeed  to execute the traininng.

### Training Data

* [coco-karpathy-opus-de dataset](https://huggingface.co/datasets/Jotschi/coco-karpathy-opus-de)

### Training Procedure 

The model was trained using PEFT 4Bit Q-LoRA with the following parameters:

* rank: 256
* alpha: 16
* steps: 8500
* bf16: True
* lr_scheduler_type: cosine
* warmup_ratio: 0.03
* gradient accumulation steps: 2
* batch size: 4
* Input sequence length: 512
* Learning Rate: 2.0e-5

#### Postprocessing

The merged model was saved using `PeftModel` API.

### Framework versions

- PEFT 0.8.2