Hugging Face's logo

ariel-ml
/
PULI-LlumiX-32K-instruct-GGUF

Text Generation
GGUF
Hungarian
English
puli
llama
finetuned
Model card Files Community
1
PULI-LlumiX-32K-instruct-GGUF
File size: 2,086 Bytes 
65d9584
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
 
---
license: llama2
language:
- hu
- en
tags:
- puli
- llama
- finetuned
base_model: ariel-ml/PULI-LlumiX-32K-instruct-f16-0.2
pipeline_tag: text-generation
---

# PULI LlumiX 32K instruct (6.74B billion parameter)

<img src="logo.webp" width="340" style="margin-left:'auto' margin-right:'auto' display:'block'"/>

Intruct finetuned version of NYTK/PULI-LlumiX-32K.

## Provided files
| Quant method | Bits | Use case |
| ---- | ---- | ---- |
| Q3_K_M | 3 | very small, high quality loss |
| Q4_K_S | 4 | small, greater quality loss |
| Q4_K_M | 4 | medium, balanced quality - recommended |
| Q5_K_S | 5 | large, low quality loss - recommended |
| Q5_K_M | 5 | large, very low quality loss - recommended |
| Q6_K | 6 | very large, extremely low quality loss |
| Q8_0 | 8 | very large, extremely low quality loss - not recommended |

## Training platform
[Runpod](https://runpod.ui) RTX 4090 GPU

## Hyper parameters

- Epoch: 3
- LoRA rank (r): 16
- LoRA alpha: 16
- Lr: 2e-4
- Lr scheduler: cosine
- Optimizer: adamw_8bit
- Weight decay: 0.01

## Dataset

boapps/szurkemarha

Only Hungarian instructions were selected: ~53000 prompts.

## Prompt format: ChatML

```
<|im_start|>system
Egy segítőkész mesterséges intelligencia asszisztens vagy. Válaszold meg a kérdést legjobb tudásod szerint!<|im_end|>
<|im_start|>user
Ki a legerősebb szuperhős?<|im_end|>
<|im_start|>assistant
A legerősebb szuperhős a Marvel univerzumában Hulk.<|im_end|>
```

## Base model

- Trained with OpenChatKit [github](https://github.com/togethercomputer/OpenChatKit)
- The [LLaMA-2-7B-32K](https://huggingface.co/togethercomputer/LLaMA-2-7B-32K) model were continuously pretrained on Hungarian dataset
- The model has been extended to a context length of 32K with position interpolation
- Checkpoint: 100 000 steps

## Base model dataset for continued pretraining

- Hungarian: 7.9 billion words, documents (763K) that exceed 5000 words in length
- English: Long Context QA (2 billion words), BookSum (78 million words)

## Limitations

- max_seq_length = 32 768
- float16
- vocab size: 32 000