winglian commited on
Commit
ae2ee9d
1 Parent(s): a713ab6

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +81 -0
README.md ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ datasets:
3
+ - anon8231489123/ShareGPT_Vicuna_unfiltered
4
+ - ehartford/wizard_vicuna_70k_unfiltered
5
+ - ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered
6
+ - QingyiSi/Alpaca-CoT
7
+ - teknium/GPT4-LLM-Cleaned
8
+ - teknium/GPTeacher-General-Instruct
9
+ - metaeval/ScienceQA_text_only
10
+ - hellaswag
11
+ - openai/summarize_from_feedback
12
+ - riddle_sense
13
+ - gsm8k
14
+ - ewof/code-alpaca-instruct-unfiltered
15
+ language:
16
+ - en
17
+ library_name: transformers
18
+ pipeline_tag: text-generation
19
+ ---
20
+
21
+ # Manticore 30B Chat (ALPHA)
22
+
23
+ - Alpha release of checkpoint before train and eval loss spikes. Additionally, there seems to be some alignment which is easily jailbroken.
24
+
25
+ Manticore 30B Chat builds on Manticore v1 with new datasets, including a de-duped subset of the Pygmalion dataset. It also removes all Alpaca style prompts using `###` in favor of
26
+ chat only style prompts using `USER:`,`ASSISTANT:` as well as [pygmalion/metharme prompting](https://huggingface.co/PygmalionAI/metharme-7b#prompting) using `<|system|>, <|user|> and <|model|>` tokens.
27
+
28
+ Questions, comments, feedback, looking to donate, or want to help? Reach out on our [Discord](https://discord.gg/EqrvvehG) or email [wing@openaccessaicollective.org](mailto:wing@openaccessaicollective.org)
29
+
30
+ # Training Datasets
31
+
32
+ Manticore 30B Chat is a Llama 30B model fine-tuned on the following datasets along with the datasets from the original Manticore 30B.
33
+
34
+ **Manticore 30B Chat was trained on effectively 40% of the datasets below due to only training for 0.4 epochs.
35
+
36
+ - de-duped pygmalion dataset, filtered down to RP data
37
+ - [riddle_sense](https://huggingface.co/datasets/riddle_sense) - instruct augmented
38
+ - hellaswag, updated for detailed explanations w 30K+ rows
39
+ - [gsm8k](https://huggingface.co/datasets/gsm8k) - instruct augmented
40
+ - [ewof/code-alpaca-instruct-unfiltered](https://huggingface.co/datasets/ewof/code-alpaca-instruct-unfiltered)
41
+
42
+ Manticore 30B
43
+ - [ShareGPT](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered) - based on a cleaned and de-suped subset
44
+ - [WizardLM](https://huggingface.co/datasets/ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered)
45
+ - [Wizard-Vicuna](https://huggingface.co/datasets/ehartford/wizard_vicuna_70k_unfiltered)
46
+ - [subset of QingyiSi/Alpaca-CoT for roleplay and CoT](https://huggingface.co/QingyiSi/Alpaca-CoT)
47
+ - [GPT4-LLM-Cleaned](https://huggingface.co/datasets/teknium/GPT4-LLM-Cleaned)
48
+ - [GPTeacher-General-Instruct](https://huggingface.co/datasets/teknium/GPTeacher-General-Instruct)
49
+ - ARC-Easy & ARC-Challenge - instruct augmented for detailed responses, derived from the `train` split
50
+ - [hellaswag](https://huggingface.co/datasets/hellaswag) - 5K row subset of instruct augmented for concise responses, derived from the `train` split
51
+ - [metaeval/ScienceQA_text_only](https://huggingface.co/datasets/metaeval/ScienceQA_text_only) - instruct for concise responses
52
+ - [openai/summarize_from_feedback](https://huggingface.co/datasets/openai/summarize_from_feedback) - instruct augmented tl;dr summarization
53
+
54
+ Not added from Manticore 13B:
55
+ - mmlu - mmlu datasets were not added to this model as the `test` split is used for benchmarks
56
+
57
+ # Shoutouts
58
+
59
+ Special thanks to Nanobit for helping with Axolotl, TheBloke for quantizing these models are more accessible to all, ehartford for cleaned datasets, and 0x000011b for the RP dataset.
60
+ # Demo
61
+
62
+ Try out the model in HF Spaces. The demo uses a quantized GGML version of the model to quickly return predictions on smaller GPUs (and even CPUs). Quantized GGML may have some minimal loss of model quality.
63
+ - https://huggingface.co/spaces/openaccess-ai-collective/manticore-13b-chat-pyg
64
+
65
+ ## Release Notes
66
+
67
+ - https://wandb.ai/wing-lian/manticore-13b-v2/runs/ij10c6m3
68
+
69
+ ## Build
70
+
71
+ Manticore was built with [Axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) on 8xA100 80GB
72
+ - 0.4 epochs taking approximately 14 hours. No further epochs will be released for the alpha.
73
+ - The configuration to duplicate this build is provided in this repo's [/config folder](https://huggingface.co/openaccess-ai-collective/manticore-30b-chat-pyg-alpha/tree/main/configs).
74
+
75
+ ## Bias, Risks, and Limitations
76
+ Manticore has not been aligned to human preferences with techniques like RLHF or deployed with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so).
77
+ Manticore was fine-tuned from the base model LlaMa 13B, please refer to its model card's Limitations Section for relevant information.
78
+
79
+ ## Examples
80
+
81
+ TBD