Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,64 @@
|
|
| 1 |
---
|
| 2 |
license: llama2
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: llama2
|
| 3 |
---
|
| 4 |
+
EXL2 quant of alpindale/goliath-120b (https://huggingface.co/alpindale/goliath-120b), to be used on exllamav2. 4.25bpw to being to able to use CFG comfortably on 72GB VRAM.
|
| 5 |
+
|
| 6 |
+
Calibration dataset is a cleaned, fixed pippa RP dataset, which does affect the results (in favor) for RP usage.
|
| 7 |
+
|
| 8 |
+
You can find the calibration dataset [here](https://discord.com/channels/1111983596572520458/1152699950208139415/1152700764230271076) (You will need to be on TheBloke server)
|
| 9 |
+
|
| 10 |
+
# Original model card
|
| 11 |
+
|
| 12 |
+
# Goliath 120B
|
| 13 |
+
|
| 14 |
+
An auto-regressive causal LM created by combining 2x finetuned [Llama-2 70B](https://huggingface.co/meta-llama/llama-2-70b-hf) into one.
|
| 15 |
+
|
| 16 |
+
Please check out the quantized formats provided by [@TheBloke](https:///huggingface.co/TheBloke) and [@Panchovix](https://huggingface.co/Panchovix):
|
| 17 |
+
|
| 18 |
+
- [GGUF](https://huggingface.co/TheBloke/goliath-120b-GGUF) (llama.cpp)
|
| 19 |
+
- [GPTQ](https://huggingface.co/TheBloke/goliath-120b-GPTQ) (KoboldAI, TGW, Aphrodite)
|
| 20 |
+
- [AWQ](https://huggingface.co/TheBloke/goliath-120b-AWQ) (TGW, Aphrodite, vLLM)
|
| 21 |
+
- [Exllamav2](https://huggingface.co/Panchovix/goliath-120b-exl2) (TGW, KoboldAI)
|
| 22 |
+
|
| 23 |
+
# Prompting Format
|
| 24 |
+
|
| 25 |
+
Both Vicuna and Alpaca will work, but due the initial and final layers belonging primarily to Xwin, I expect Vicuna to work the best.
|
| 26 |
+
|
| 27 |
+
# Merge process
|
| 28 |
+
|
| 29 |
+
The models used in the merge are [Xwin](https://huggingface.co/Xwin-LM/Xwin-LM-70B-V0.1) and [Euryale](https://huggingface.co/Sao10K/Euryale-1.3-L2-70B).
|
| 30 |
+
|
| 31 |
+
The layer ranges used are as follows:
|
| 32 |
+
|
| 33 |
+
```yaml
|
| 34 |
+
- range 0, 16
|
| 35 |
+
Xwin
|
| 36 |
+
- range 8, 24
|
| 37 |
+
Euryale
|
| 38 |
+
- range 17, 32
|
| 39 |
+
Xwin
|
| 40 |
+
- range 25, 40
|
| 41 |
+
Euryale
|
| 42 |
+
- range 33, 48
|
| 43 |
+
Xwin
|
| 44 |
+
- range 41, 56
|
| 45 |
+
Euryale
|
| 46 |
+
- range 49, 64
|
| 47 |
+
Xwin
|
| 48 |
+
- range 57, 72
|
| 49 |
+
Euryale
|
| 50 |
+
- range 65, 80
|
| 51 |
+
Xwin
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
# Screenshots
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
# Benchmarks
|
| 59 |
+
Coming soon.
|
| 60 |
+
|
| 61 |
+
# Acknowledgements
|
| 62 |
+
Credits goes to [@chargoddard](https://huggingface.co/chargoddard) for developing the framework used to merge the model - [mergekit](https://github.com/cg123/mergekit).
|
| 63 |
+
|
| 64 |
+
Special thanks to [@Undi95](https://huggingface.co/Undi95) for helping with the merge ratios.
|