Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: llama2
|
| 3 |
+
language:
|
| 4 |
+
- hu
|
| 5 |
+
- en
|
| 6 |
+
tags:
|
| 7 |
+
- puli
|
| 8 |
+
- llama
|
| 9 |
+
- finetuned
|
| 10 |
+
base_model: ariel-ml/PULI-LlumiX-32K-instruct-f16-0.2
|
| 11 |
+
pipeline_tag: text-generation
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# PULI LlumiX 32K instruct (6.74B billion parameter)
|
| 15 |
+
|
| 16 |
+
<img src="logo.webp" width="340" style="margin-left:'auto' margin-right:'auto' display:'block'"/>
|
| 17 |
+
|
| 18 |
+
Intruct finetuned version of NYTK/PULI-LlumiX-32K.
|
| 19 |
+
|
| 20 |
+
## Provided files
|
| 21 |
+
| Quant method | Bits | Use case |
|
| 22 |
+
| ---- | ---- | ---- |
|
| 23 |
+
| Q3_K_M | 3 | very small, high quality loss |
|
| 24 |
+
| Q4_K_S | 4 | small, greater quality loss |
|
| 25 |
+
| Q4_K_M | 4 | medium, balanced quality - recommended |
|
| 26 |
+
| Q5_K_S | 5 | large, low quality loss - recommended |
|
| 27 |
+
| Q5_K_M | 5 | large, very low quality loss - recommended |
|
| 28 |
+
| Q6_K | 6 | very large, extremely low quality loss |
|
| 29 |
+
| Q8_0 | 8 | very large, extremely low quality loss - not recommended |
|
| 30 |
+
|
| 31 |
+
## Training platform
|
| 32 |
+
[Runpod](https://runpod.ui) RTX 4090 GPU
|
| 33 |
+
|
| 34 |
+
## Hyper parameters
|
| 35 |
+
|
| 36 |
+
- Epoch: 3
|
| 37 |
+
- LoRA rank (r): 16
|
| 38 |
+
- LoRA alpha: 16
|
| 39 |
+
- Lr: 2e-4
|
| 40 |
+
- Lr scheduler: cosine
|
| 41 |
+
- Optimizer: adamw_8bit
|
| 42 |
+
- Weight decay: 0.01
|
| 43 |
+
|
| 44 |
+
## Dataset
|
| 45 |
+
|
| 46 |
+
boapps/szurkemarha
|
| 47 |
+
|
| 48 |
+
Only Hungarian instructions were selected: ~53000 prompts.
|
| 49 |
+
|
| 50 |
+
## Prompt format: ChatML
|
| 51 |
+
|
| 52 |
+
```
|
| 53 |
+
<|im_start|>system
|
| 54 |
+
Egy segítőkész mesterséges intelligencia asszisztens vagy. Válaszold meg a kérdést legjobb tudásod szerint!<|im_end|>
|
| 55 |
+
<|im_start|>user
|
| 56 |
+
Ki a legerősebb szuperhős?<|im_end|>
|
| 57 |
+
<|im_start|>assistant
|
| 58 |
+
A legerősebb szuperhős a Marvel univerzumában Hulk.<|im_end|>
|
| 59 |
+
```
|
| 60 |
+
|
| 61 |
+
## Base model
|
| 62 |
+
|
| 63 |
+
- Trained with OpenChatKit [github](https://github.com/togethercomputer/OpenChatKit)
|
| 64 |
+
- The [LLaMA-2-7B-32K](https://huggingface.co/togethercomputer/LLaMA-2-7B-32K) model were continuously pretrained on Hungarian dataset
|
| 65 |
+
- The model has been extended to a context length of 32K with position interpolation
|
| 66 |
+
- Checkpoint: 100 000 steps
|
| 67 |
+
|
| 68 |
+
## Base model dataset for continued pretraining
|
| 69 |
+
|
| 70 |
+
- Hungarian: 7.9 billion words, documents (763K) that exceed 5000 words in length
|
| 71 |
+
- English: Long Context QA (2 billion words), BookSum (78 million words)
|
| 72 |
+
|
| 73 |
+
## Limitations
|
| 74 |
+
|
| 75 |
+
- max_seq_length = 32 768
|
| 76 |
+
- float16
|
| 77 |
+
- vocab size: 32 000
|