README.md

---
language:
- sv
license: llama3
library_name: transformers
tags:
- unsloth
datasets:
- neph1/bellman-7b-finetune
- neph1/codefeedback-swedish
---

# Model Card for Bellman

This version of bellman is finetuned from llama-3-instruct-8b.
It is arguable whether it's better at Swedish, because llama-3 is really good. It's however finetuned for prompt question answering, based on a dataset created from 
Swedish wikipedia, with a lot of Sweden-centric questions.
New from previous versions is questions from a translated code-feedback dataset, as well as a number of stories. It's not great at generating stories, 
but better than previosly.

Please note, the HuggingFace inference api is probably trying to load the adapter (lora) which isn't going to work.

240609: I've uploaded a 4-bit GPTQ quant, but it's completely untested.

![image/png](https://cdn-uploads.huggingface.co/production/uploads/653cd3049107029eb004f968/IDGX3d9lGe6yx-yHjsrav.png)

## Model Details

Training run on 240606:

Step 	Training Loss 	Validation Loss<br>
 25 	1.506400 	1.164538<br>
 50 	1.128500 	1.059316<br>
 75 	1.095100 	1.040511<br>
100 	1.068700 	1.031033<br>
125 	1.061300 	1.024377<br>
150 	1.035700 	1.017490<br>
175 	1.061200 	1.012095<br>
200 	1.031600 	1.007867<br>
225 	1.031900 	1.002652<br>
250 	0.958300 	1.003817<br>
275 	0.967900 	1.000483<br>
300 	0.950000 	0.998807<br>
325 	0.974300 	0.996894<br>
350 	0.960700 	0.994098<br>
375 	0.956000 	0.991491<br>
400 	0.940500 	0.988697<br>
425 	0.949100 	0.987253<br>
450 	0.940600 	0.986425 <-- Picked checkpoint<br>
475 	0.888300 	0.994204<br>
500 	0.881700 	0.994897<br>

### Model Description


- **Developed by:** Me
- **Funded by:** Me
- **Model type:** Instruct
- **Language(s) (NLP):** Swedish
- **License:** llama-3
- **Finetuned from model:** Llama3 Instruct 8b

## Model Card Contact

[email protected]