File size: 963 Bytes
50cc00d
 
52eebcc
50cc00d
 
 
80257a6
50cc00d
59694f4
 
 
 
 
 
 
 
 
 
b78f32c
afa0303
80257a6
50cc00d
80257a6
fe0d983
 
4ff7825
fe0d983
 
 
 
 
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
---
datasets:
- Aeala/ShareGPT_Vicuna_unfiltered
---

## LoRA Info:
Please note that this is a highly experimental LoRA model. It may do some good stuff, it might do some undesirable stuff. Training is basically done now. Feel free to try it!~

**Important Note**: While this is trained on a cleaned ShareGPT dataset like Vicuna used, this was trained in the *Alpaca* format, so prompting should be something like:

```
### Instruction:

<prompt> (without the <>)

### Response:
```

Current upload: Fully trained adapter model (3 epochs).

Secondary upload: checkpoint of epoch 2.97 (of 3)

Thanks to MetaIX for initial seemingly successful testing of the first uploaded checkpoint (epoch 0.8) as well as epoch 1.
## Benchmarks
**wikitext2:** 4.372413635253906

**ptb-new:** 24.69171714782715

**c4-new:** 6.469308853149414

Results generated with GPTQ evals (not quantized) thanks to [Neko-Institute-of-Science](https://huggingface.co/Neko-Institute-of-Science)