Decoding Strategies in Large Language Models
•
38
Maxime Labonne PRO
mlabonne
AI & ML interests
Post-training, model editing, quantization
Recent Activity
liked
a Space
2 days ago
burtenshaw/recap
Articles
Organizations
Posts
5
Post
17285
Large models are surprisingly bad storytellers.
I asked 8 LLMs to "Tell me a bedtime story about bears and waffles."
Claude 3.5 Sonnet and GPT-4o gave me the worst stories: no conflict, no moral, zero creativity.
In contrast, smaller models were quite creative and wrote stories involving talking waffle trees and bears ostracized for their love of waffles.
Here you can see a comparison between Claude 3.5 Sonnet and NeuralDaredevil-8B-abliterated. They both start with a family of bears but quickly diverge in terms of personality, conflict, etc.
I mapped it to the hero's journey to have some kind of framework. Prompt engineering can definitely help here, but it's still disappointing that the larger models don't create better stories right off the bat.
Do you know why smaller models outperform the frontier models here?
I asked 8 LLMs to "Tell me a bedtime story about bears and waffles."
Claude 3.5 Sonnet and GPT-4o gave me the worst stories: no conflict, no moral, zero creativity.
In contrast, smaller models were quite creative and wrote stories involving talking waffle trees and bears ostracized for their love of waffles.
Here you can see a comparison between Claude 3.5 Sonnet and NeuralDaredevil-8B-abliterated. They both start with a family of bears but quickly diverge in terms of personality, conflict, etc.
I mapped it to the hero's journey to have some kind of framework. Prompt engineering can definitely help here, but it's still disappointing that the larger models don't create better stories right off the bat.
Do you know why smaller models outperform the frontier models here?
Post
17847
✂️ Uncensor any LLM with abliteration
I wrote an article about abliteration and how NeuralDaredevil-8B was created. Beyond removing alignment, I believe it's an interesting technique with a lot of potential. It's basically fine-tuning without retraining.
In this article, we see how it works, implement it in Google Colab, and heal the abliterated model to recover the performance drop due to this technique. The final model is an uncensored and high-quality model with the highest MMLU score on the Open LLM Leaderboard (8B category).
https://huggingface.co./blog/mlabonne/abliteration
I wrote an article about abliteration and how NeuralDaredevil-8B was created. Beyond removing alignment, I believe it's an interesting technique with a lot of potential. It's basically fine-tuning without retraining.
In this article, we see how it works, implement it in Google Colab, and heal the abliterated model to recover the performance drop due to this technique. The final model is an uncensored and high-quality model with the highest MMLU score on the Open LLM Leaderboard (8B category).
https://huggingface.co./blog/mlabonne/abliteration
Collections
10
Uncensored models using abliteration. See this article for more information: huggingface.co/blog/mlabonne/abliteration
-
mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated
Text Generation • Updated • 155k • 141 -
mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated-GGUF
Updated • 15.4k • 127 -
mlabonne/Llama-3.1-70B-Instruct-lorablated
Text Generation • Updated • 2.65k • 65 -
mlabonne/Llama-3.1-70B-Instruct-lorablated-GGUF
Updated • 2.29k • 41
Fine-tuned abliterated merge of the best Llama 3 8B model. Highest MMLU score in its category.
spaces
19
models
145
mlabonne/codellama-2-7b
Updated
•
65
•
4
mlabonne/dummy-llama-2
Text Generation
•
Updated
•
1.06k
•
8
mlabonne/Hermes-3-Llama-3.1-70B-lorablated
Text Generation
•
Updated
•
739
•
25
mlabonne/TwinLlama-3.1-8B-DPO-GGUF
Updated
•
21
•
2
mlabonne/TwinLlama-3.1-8B-DPO
Text Generation
•
Updated
•
138
•
3
mlabonne/TwinLlama-3.1-8B-GGUF
Updated
•
24
•
2
mlabonne/TwinLlama-3.1-8B
Text Generation
•
Updated
•
174
•
8
mlabonne/UltraLlama-3.1-8B
Updated
mlabonne/BigQwen2.5-Echo-47B-Instruct
Text Generation
•
Updated
•
24
•
3
mlabonne/BigQwen2.5-52B-Instruct
Text Generation
•
Updated
•
44
•
4
datasets
48
mlabonne/smoltalk-flat
Viewer
•
Updated
•
1.1M
•
240
•
4
mlabonne/orca-agentinstruct-1M-v1-cleaned
Viewer
•
Updated
•
1.05M
•
1.76k
•
52
mlabonne/open-perfectblend
Viewer
•
Updated
•
1.42M
•
144
•
45
mlabonne/lmsys-arena-human-preference-55k-sharegpt
Viewer
•
Updated
•
57.4k
•
1.45k
•
7
mlabonne/orpo-dpo-mix-40k
Viewer
•
Updated
•
44.2k
•
1.4k
•
262
mlabonne/ultrachat_200k_sft
Viewer
•
Updated
•
208k
•
251
•
1
mlabonne/orca-math-word-problems-80k
Viewer
•
Updated
•
80k
•
73
•
4
mlabonne/llmtwin-dpo
Viewer
•
Updated
•
1.63k
•
89
•
1
mlabonne/llmtwin
Viewer
•
Updated
•
3.34k
•
123
•
7
mlabonne/lmsys-arena-human-preference-filtered-19k
Viewer
•
Updated
•
57.5k
•
50
•
4