Merging in a Bottle: Differentiable Adaptive Merging (DAM) and the Path from Averaging to Automation
Paper
•
2410.08371
•
Published
•
1
Liquid AI
Enterprise
company
AI & ML interests
A new generation of foundation models from first principles.
Recent Activity
LiquidAI's activity
mlabonne
authored
a
paper
about 2 months ago
jimmysmith1919
authored
3
papers
2 months ago
Post
18211
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?
- 44 replies
Post
18825
✂️ 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
-
- 26 replies
Post
2177
FiftyOne Datasets <> Hugging Face Hub Integration!
As of yesterday's release of FiftyOne
You can now load Parquet datasets from the hub and have them converted directly into FiftyOne datasets. To load MNIST, for example:
You can also load FiftyOne datasets directly from the hub. Here's how you load the first 1000 samples from the VisDrone dataset:
And tying it all together, you can push your FiftyOne datasets directly to the hub:
Major thanks to @tomaarsen @davanstrien @severo @osanseviero and @julien-c for helping to make this happen!!!
Full documentation and details here: https://docs.voxel51.com/integrations/huggingface.html#huggingface-hub
As of yesterday's release of FiftyOne
0.23.8, the FiftyOne open source library for dataset curation and visualization is now integrated with the Hugging Face Hub!You can now load Parquet datasets from the hub and have them converted directly into FiftyOne datasets. To load MNIST, for example:
pip install -U fiftyoneimport fiftyone as fo
import fiftyone.utils.huggingface as fouh
dataset = fouh.load_from_hub(
"mnist",
format="ParquetFilesDataset",
classification_fields="label",
)
session = fo.launch_app(dataset)You can also load FiftyOne datasets directly from the hub. Here's how you load the first 1000 samples from the VisDrone dataset:
import fiftyone as fo
import fiftyone.utils.huggingface as fouh
dataset = fouh.load_from_hub("jamarks/VisDrone2019-DET", max_samples=1000)
# Launch the App
session = fo.launch_app(dataset)And tying it all together, you can push your FiftyOne datasets directly to the hub:
import fiftyone.zoo as foz
import fiftyone.utils.huggingface as fouh
dataset = foz.load_zoo_dataset("quickstart")
fouh.push_to_hub(dataset, "my-dataset")Major thanks to @tomaarsen @davanstrien @severo @osanseviero and @julien-c for helping to make this happen!!!
Full documentation and details here: https://docs.voxel51.com/integrations/huggingface.html#huggingface-hub
- 3 replies
Post
12784
🔁 AutoMerger created the best 7B model on the Open LLM Leaderboard
By randomly combining top models from the Open LLM Leaderboard, AutoMerger created YamshadowExperiment28-7B. The model is three weeks old and has been at the top of the leaderboard for a week now. It was created through a simple SLERP merge of:
- automerger/YamShadow-7B (another top model created by AutoMerger)
- yam-peleg/Experiment28-7B (a top model from @yam-peleg )
1/ On the Open LLM Leaderboard, it managed to outperform the excellent M7-7b model, which has been the #1 7B model for a while now.
2/ On the YALL leaderboard, YamshadowExperiment28-7B is ranked as the 9th best-performing automerge (but note that the scores are very close to each other). Compared to others, it does not perform particularly well on AGIEval or Bigbench.
3/ Thanks to @sam-paech , I have scores on EQ-Bench, where it managed to outperform all of my previous models. It even surpasses recent models such as DBRX instruct, Qwen1.5 32B Chat, and Cohere's Command R+.
Surprisingly, it does not support ChatML or Mistral Instruct, unlike my other merges (which are part of its family tree). Alpaca works well 99% of the time, but the model can sometimes produce a lot of "INST" tokens for no reason.
In my experiments, YamshadowExperiment28-7B doesn't seem smarter than other successful merges like AlphaMonarch. On the contrary, I found several mathematical or reasoning problems where it fails.
Considering these results, it looks like it might overfit the Open LLM Leaderboard. I guess it's anything but surprising when you randomly merge 156 models.
🤗 Model: automerger/YamshadowExperiment28-7B
🔁 AutoMerger: mlabonne/AutoMerger
By randomly combining top models from the Open LLM Leaderboard, AutoMerger created YamshadowExperiment28-7B. The model is three weeks old and has been at the top of the leaderboard for a week now. It was created through a simple SLERP merge of:
- automerger/YamShadow-7B (another top model created by AutoMerger)
- yam-peleg/Experiment28-7B (a top model from @yam-peleg )
1/ On the Open LLM Leaderboard, it managed to outperform the excellent M7-7b model, which has been the #1 7B model for a while now.
2/ On the YALL leaderboard, YamshadowExperiment28-7B is ranked as the 9th best-performing automerge (but note that the scores are very close to each other). Compared to others, it does not perform particularly well on AGIEval or Bigbench.
3/ Thanks to @sam-paech , I have scores on EQ-Bench, where it managed to outperform all of my previous models. It even surpasses recent models such as DBRX instruct, Qwen1.5 32B Chat, and Cohere's Command R+.
Surprisingly, it does not support ChatML or Mistral Instruct, unlike my other merges (which are part of its family tree). Alpaca works well 99% of the time, but the model can sometimes produce a lot of "INST" tokens for no reason.
In my experiments, YamshadowExperiment28-7B doesn't seem smarter than other successful merges like AlphaMonarch. On the contrary, I found several mathematical or reasoning problems where it fails.
Considering these results, it looks like it might overfit the Open LLM Leaderboard. I guess it's anything but surprising when you randomly merge 156 models.
🤗 Model: automerger/YamshadowExperiment28-7B
🔁 AutoMerger: mlabonne/AutoMerger
Post
9266
⚡ AutoQuant
AutoQuant is the evolution of my previous AutoGGUF notebook (https://colab.research.google.com/drive/1P646NEg33BZy4BfLDNpTz0V0lwIU3CHu). It allows you to quantize your models in five different formats:
- GGUF: perfect for inference on CPUs (and LM Studio)
- GPTQ/EXL2: fast inference on GPUs
- AWQ: super fast inference on GPUs with vLLM (https://github.com/vllm-project/vllm)
- HQQ: extreme quantization with decent 2-bit and 3-bit models
Once the model is converted, it automatically uploads it on the Hugging Face Hub. To quantize a 7B model, GGUF only needs a T4 GPU, while the other methods require an A100 GPU.
Here's an example of a model I quantized using HQQ and AutoQuant: mlabonne/AlphaMonarch-7B-2bit-HQQ
I hope you'll enjoy it and quantize lots of models! :)
💻 AutoQuant: https://colab.research.google.com/drive/1b6nqC7UZVt8bx4MksX7s656GXPM-eWw4
AutoQuant is the evolution of my previous AutoGGUF notebook (https://colab.research.google.com/drive/1P646NEg33BZy4BfLDNpTz0V0lwIU3CHu). It allows you to quantize your models in five different formats:
- GGUF: perfect for inference on CPUs (and LM Studio)
- GPTQ/EXL2: fast inference on GPUs
- AWQ: super fast inference on GPUs with vLLM (https://github.com/vllm-project/vllm)
- HQQ: extreme quantization with decent 2-bit and 3-bit models
Once the model is converted, it automatically uploads it on the Hugging Face Hub. To quantize a 7B model, GGUF only needs a T4 GPU, while the other methods require an A100 GPU.
Here's an example of a model I quantized using HQQ and AutoQuant: mlabonne/AlphaMonarch-7B-2bit-HQQ
I hope you'll enjoy it and quantize lots of models! :)
💻 AutoQuant: https://colab.research.google.com/drive/1b6nqC7UZVt8bx4MksX7s656GXPM-eWw4
-
-
- 16 replies
Post
🌳 Model Family Tree
Merging models has become a powerful way to compress information and build powerful models for cheap. Right now, the process is still quite experimental: which models to merge? which parameters should I use? We have some intuition but no principled approach.
I made a little tool to make things a little clearer. It allows you to visualize the family tree of any model on the Hub. It also displays the type of license they use: permissive (green), noncommercial (red), and unknown (gray). It should help people select the right license based on the parent models.
In addition, I hope it can be refined to extract more information about these models: do models from very different branches work better when merged? Can we select them based on the weight difference? There are a lot of questions to explore in this new space. :)
Here's a link to the colab notebook I made: https://colab.research.google.com/drive/1s2eQlolcI1VGgDhqWIANfkfKvcKrMyNr
If you want to know more about model merging or build you own merges, here's the article I wrote about this topic: https://huggingface.co/blog/mlabonne/merge-models
Merging models has become a powerful way to compress information and build powerful models for cheap. Right now, the process is still quite experimental: which models to merge? which parameters should I use? We have some intuition but no principled approach.
I made a little tool to make things a little clearer. It allows you to visualize the family tree of any model on the Hub. It also displays the type of license they use: permissive (green), noncommercial (red), and unknown (gray). It should help people select the right license based on the parent models.
In addition, I hope it can be refined to extract more information about these models: do models from very different branches work better when merged? Can we select them based on the weight difference? There are a lot of questions to explore in this new space. :)
Here's a link to the colab notebook I made: https://colab.research.google.com/drive/1s2eQlolcI1VGgDhqWIANfkfKvcKrMyNr
If you want to know more about model merging or build you own merges, here's the article I wrote about this topic: https://huggingface.co/blog/mlabonne/merge-models
- 8 replies
Zymrael
authored
8
papers
about 1 year ago
Hyena Hierarchy: Towards Larger Convolutional Language Models
Paper
•
2302.10866
•
Published
•
7
Deep Latent State Space Models for Time-Series Generation
Paper
•
2212.12749
•
Published
•
1
Neural Solvers for Fast and Accurate Numerical Optimal Control
Paper
•
2203.08072
•
Published
Transform Once: Efficient Operator Learning in Frequency Domain
Paper
•
2211.14453
•
Published
Effectively Modeling Time Series with Simple Discrete State Spaces
Paper
•
2303.09489
•
Published
•
1
Monarch Mixer: A Simple Sub-Quadratic GEMM-Based Architecture
Paper
•
2310.12109
•
Published
•
1
Monarch: Expressive Structured Matrices for Efficient and Accurate Training
Paper
•
2204.00595
•
Published
•
1
Laughing Hyena Distillery: Extracting Compact Recurrences From Convolutions
Paper
•
2310.18780
•
Published
•
3
alexanderamini
authored
a
paper
about 1 year ago
haerski
authored
a
paper
over 1 year ago