adamo1139 commited on
Commit
cb00ce7
1 Parent(s): 455b1a9

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +50 -0
README.md CHANGED
@@ -2,4 +2,54 @@
2
  license: other
3
  license_name: other
4
  license_link: LICENSE
 
 
 
 
5
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  license: other
3
  license_name: other
4
  license_link: LICENSE
5
+ datasets:
6
+ - adamo1139/rawrr_v2
7
+ - adamo1139/AEZAKMI_v3-6
8
+ - unalignment/toxic-dpo-v0.1
9
  ---
10
+ ## Model description
11
+
12
+ Yi-34B 200K XLCTX base model fine-tuned on RAWrr_v2 (DPO), AEZAKMI-3-6 (SFT) and unalignment/toxic-dpo-0.1 (DPO) datasets. Training took around 20-30 hours total on RTX 3090 Ti, all finetuning was done locally.
13
+ It's like airoboros but with less gptslop, no refusals and less typical language used by RLHFed OpenAI models, with extra spicyness.
14
+ Say goodbye to "It's important to remember"! \
15
+ Prompt format is standard chatml. Don't expect it to be good at math, riddles or be crazy smart. My end goal with AEZAKMI is to create a cozy free chatbot.
16
+ Cost of this fine-tune is about $5-$10 in electricity.
17
+ Base model used for fine-tuning was Yi-34B-200K model shared by 01.ai, the newer version that has improved long context needle in a haystack retrieval. They didn't give it a new name, giving it numbers would mess up AEZAKMI naming scheme by adding a second number, so I will be calling it XLCTX.
18
+
19
+
20
+ I had to lower max_positional_embeddings in config.json and model_max_length for training to start, otherwise I was OOMing straight away.
21
+ This attempt had both max_position_embeddings and model_max_length set to 4096, which worked perfectly fine. I then reversed this to 200000 once I was uploading it.
22
+ I think it should keep long context capabilities of the base model.
23
+
24
+ ## Quants!
25
+
26
+ EXL2 quants coming soon, I think I will start by uploading 4bpw quant in a few days.
27
+
28
+
29
+ ## Prompt Format
30
+
31
+ I recommend using ChatML format, as this was used during fine-tune. \
32
+ Here's a prompt format you should use, you can set a different system message, model was trained on SystemChat dataset, so it should respect system prompts fine.
33
+
34
+ ```
35
+ <|im_start|>system
36
+ A chat.<|im_end|>
37
+ <|im_start|>user
38
+ {prompt}<|im_end|>
39
+ <|im_start|>assistant
40
+ ```
41
+
42
+ ## Intended uses & limitations
43
+
44
+ Use is limited by Yi license.
45
+
46
+ ## Known Issues
47
+
48
+ It's more of an assistant feel rather than a human feel, at least with system chat "A chat." \
49
+ Long context wasn't tested yet, it should work fine though - feel free to give me feedback about it.
50
+
51
+ ## Credits
52
+
53
+ Thanks to unsloth and huggingface team for providing software packages used during fine-tuning. \
54
+ Thanks to Jon Durbin, abacusai, huggingface, sandex, NobodyExistsOnTheInternet, Nous-Research for open sourcing datasets I included in the AEZAKMI dataset. \
55
+ AEZAKMI is basically a mix of open source datasets I found on HF, so without them this would not be possible at all.