lucyknada commited on
Commit
f791c02
1 Parent(s): 2857ff5

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +65 -0
README.md ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ - fr
6
+ - de
7
+ - es
8
+ - it
9
+ - pt
10
+ - ru
11
+ - zh
12
+ - ja
13
+ pipeline_tag: text-generation
14
+ tags:
15
+ - chat
16
+ ---
17
+
18
+ ## This repo contains EXL2 quants of the model. If you need the original weights, please find them [here](https://huggingface.co/anthracite-org/magnum-v2-123b).
19
+ ## Base repo only contains the measurement file, see revisions for your quant of choice.
20
+
21
+ - [measurement.json](https://huggingface.co/anthracite-org/magnum-v2-123b-exl2/tree/main)
22
+ - [2.7bpw](https://huggingface.co/anthracite-org/magnum-v2-123b-exl2/tree/2.7bpw)
23
+ - [4.0bpw](https://huggingface.co/anthracite-org/magnum-v2-123b-exl2/tree/4.0bpw)
24
+ - [6.0bpw](https://huggingface.co/anthracite-org/magnum-v2-123b-exl2/tree/6.0bpw)
25
+
26
+
27
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6491e00e057b0928b3e07b75/hkPzhL-xYPeGGKCyAf3Qd.png)
28
+ This is the sixth in a series of models designed to replicate the prose quality of the Claude 3 models, specifically Sonnet and Opus. This model is fine-tuned on top of [Mistral-Large-Instruct-2407](https://huggingface.co/mistralai/Mistral-Large-Instruct-2407).
29
+
30
+ ## Prompting
31
+ Model has been Instruct tuned with the Mistral formatting. A typical input would look like this:
32
+
33
+ ```py
34
+ <s>[INST] SYSTEM MESSAGE\nUSER MESSAGE[/INST] ASSISTANT MESSAGE</s>[INST] USER MESSAGE[/INST]
35
+ ```
36
+
37
+ We also provide SillyTavern presets for [Context](https://huggingface.co/anthracite-org/Magnum-123b-v1/resolve/main/Magnum-Mistral-Context.json) and [Instruct](https://huggingface.co/anthracite-org/Magnum-123b-v1/raw/main/Magnum-Mistral-Instruct.json) respectively.
38
+
39
+ The Mistral preset included in SillyTavern seems to be misconfigured by default, so we recommend using these as a replacement.
40
+
41
+ ## Credits
42
+ - [anthracite-org/Stheno-Data-Filtered](https://huggingface.co/datasets/anthracite-org/Stheno-Data-Filtered)
43
+ - [anthracite-org/kalo-opus-instruct-22k-no-refusal](https://huggingface.co/datasets/anthracite-org/kalo-opus-instruct-22k-no-refusal)
44
+ - [anthracite-org/nopm_claude_writing_fixed](https://huggingface.co/datasets/anthracite-org/nopm_claude_writing_fixed)
45
+
46
+ This model has been a team effort, and the credits goes to all members of Anthracite.
47
+
48
+ ## Training
49
+ The training was done for 1.5 epochs. We used 8x [AMD Instinct™ MI300X Accelerators](https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html) for the full-parameter fine-tuning of the model.
50
+
51
+ In addition to this, we noticed that Mistral Large models seemed much more sensitive to learning rate adjustments than other models:
52
+
53
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6491e00e057b0928b3e07b75/xCK3ISKF6pWcMyO7MEzTA.png)
54
+
55
+ We hypothesize this is primarily due to the particularly narrow and low variance weight distributions typical of Mistral derived models regardless of their scale.
56
+
57
+ In the end, due to the costs that would be involved in training another full 2 epochs run ($600) on an even lower rate, we settled on our third attempt: 2e-6 with an effective batch size of 64, stopped earlier than the target 2 epochs.
58
+
59
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6491e00e057b0928b3e07b75/d9_cBy-DuWrdnoVBbAvRV.png)
60
+ We notice a correlation between the significance of the 2nd epoch loss drop and the strength of the learning rate, implying 4e-6 leads to more catastrophic forgetting.
61
+
62
+ [<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
63
+
64
+ ## Safety
65
+ ...