Commit
•
82cb133
1
Parent(s):
dbf308a
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,53 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model_relation: quantized
|
| 4 |
+
quantized_by: Quant-Cartel
|
| 5 |
+
base_model: rAIFle/Luca-MN-bf16
|
| 6 |
+
pipeline_tag: text-generation
|
| 7 |
+
tags:
|
| 8 |
+
- iMat
|
| 9 |
+
- GGUF
|
| 10 |
+
- unsloth
|
| 11 |
+
- trl
|
| 12 |
+
- sft
|
| 13 |
+
---
|
| 14 |
+
```
|
| 15 |
+
e88 88e d8
|
| 16 |
+
d888 888b 8888 8888 ,"Y88b 888 8e d88
|
| 17 |
+
C8888 8888D 8888 8888 "8" 888 888 88b d88888
|
| 18 |
+
Y888 888P Y888 888P ,ee 888 888 888 888
|
| 19 |
+
"88 88" "88 88" "88 888 888 888 888
|
| 20 |
+
b
|
| 21 |
+
8b,
|
| 22 |
+
|
| 23 |
+
e88'Y88 d8 888
|
| 24 |
+
d888 'Y ,"Y88b 888,8, d88 ,e e, 888
|
| 25 |
+
C8888 "8" 888 888 " d88888 d88 88b 888
|
| 26 |
+
Y888 ,d ,ee 888 888 888 888 , 888
|
| 27 |
+
"88,d88 "88 888 888 888 "YeeP" 888
|
| 28 |
+
|
| 29 |
+
PROUDLY PRESENTS
|
| 30 |
+
```
|
| 31 |
+
# Luca-MN-iMat-GGUF
|
| 32 |
+
|
| 33 |
+
Quantized with love from fp32.
|
| 34 |
+
|
| 35 |
+
* Importance Matrix calculated using [groups_merged.txt](https://github.com/ggerganov/llama.cpp/discussions/5263#discussioncomment-8395384)
|
| 36 |
+
* 92 chunks
|
| 37 |
+
* n_ctx=512
|
| 38 |
+
* Importance Matrix uses fp32 precision model weights, fp32.imatrix file to be added in repo
|
| 39 |
+
|
| 40 |
+
Original model README [here](https://huggingface.co/rAIFle/Luca-MN-bf16) and below:
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
## Luca-MN-iMat-GGUF
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
This thing was just intended as an experiment but it turned out quite good. I had it both name and prompt imagegen for itself.
|
| 48 |
+
|
| 49 |
+
Created by running a high-r LoRA-pass over Nemo-Base with 2 epochs of some RP data, then a low-r pass with 0.5 epochs of the c2-data, then 3 epochs of DPO using [jondurbin/gutenberg-dpo-v0.1](https://huggingface.co/datasets/jondurbin/gutenberg-dpo-v0.1).
|
| 50 |
+
|
| 51 |
+
## Prompting
|
| 52 |
+
|
| 53 |
+
Use the `Mistral V3-Tekken` context- and instruct-templates. Temperature at about `1.25` seems to be the sweet spot, with either MinP at `0.05` or TopP at `0.9`. DRY/Smoothing etc depending on your preference.
|