Semanlink - Favoris ; Machine learning AND Parameter-efficient fine-tuning (PEFT)

> the paper LoRA: Low-Rank Adaptation of Large Language Models proposes to decompose the weight changes, ΔW, into a lower-rank representation. (To be technically correct, LoRA does not decompose the matrices directly, but it learns the decomposed matrices via backpropagation).
>
> suppose ΔW is the weight update for an A × B weight matrix. Then, we can decompose the weight update matrix into two smaller matrices: ΔW = WA WB, where WA is an an A × r-dimensional matrix, and WB is an an r × B-dimensional matrix.

LORA and LLama :

> Lit-LLaMA repository a simple, readable reimplementation of Meta’s popular LLaMA model. Besides code for training and running LLaMA itself (with the original Meta LLaMA weights), it also contains code for finetuning LLaMA using LLaMA-Adapter and LoRA.

2023-07-27 About

What is low-rank adaptation (LoRA)? - TechTalks

Tags:

2023-07-27 About

LoRA for semantic similarity tasks

Tags:

2023-07-21 About

clem 🤗 sur Twitter : "Llama 2 by @Meta is already integrated with @huggingface transformers, TGI, inference endpoints, PEFT and much more..."

Tags:

2023-07-19 About

How to use Alpaca-LoRA to fine-tune a model like ChatGPT

Tags:

2023-05-26 About

Google "We Have No Moat, And Neither Does OpenAI"

Tags:

> low-cost public involvement was enabled
by a vastly cheaper mechanism for fine tuning called low
rank adaptation ()[LoRA](tag:lora)

> **Part of what makes LoRA so effective is that ... it’s stackable.**
>
> By contrast, training giant models from scratch not only
throws away the pretraining, but also any iterative
improvements that have been made on top.

> LoRA updates are very cheap to produce (~$100) for the
most popular model sizes.

> Many of these projects are saving time by training on
small, highly curated datasets...
> These
datasets are built using synthetic methods (e.g. filtering
the best responses from an existing model) and
scavenging from other projects

> Directly Competing With Open Source
Is a Losing Proposition

> Paradoxically, the one clear winner in all of this is Meta.
Because the leaked model was theirs ([LLaMA](tag:llama)), they have
effectively garnered an entire planet's worth of free labor.
Since most open source innovation is happening on top of
their architecture, there is nothing stopping them from
directly incorporating it into their products.

2023-05-04 About

Andrej Karpathy sur Twitter : "There's a chance that LoRA finetunes work so well that it dramatically alters the finetuning vs. retrieval + few-shot prompting power dynamic..."

Tags:

2023-04-21 About

[2106.09685] LoRA: Low-Rank Adaptation of Large Language Models

Tags:

2023-03-21 About

(((ل()(ل() 'yoav))))👾 sur Twitter : "there is this genre of papers that show you can train/tune only some subset of a network's weights, freezing the rest, and things still work as well as (or better than) full training/tuning..."

Tags:

2023-02-15 About

Parameter-Efficient Fine-Tuning using 🤗 PEFT

Tags:

Parameter-efficient fine-tuning (PEFT)

2023-02-10 About

[2002.01808] K-Adapter: Infusing Knowledge into Pre-Trained Models with Adapters

Tags:

2023-01-12 About

[2205.12410] AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning

Tags:

2022-12-16 About

[2205.05638] Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning

Tags:

2022-12-15 About

[2106.10199] BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models

Tags:

2022-09-01 About

AdapterHub: A Framework for Adapting Transformers | Towards Data Science

Tags:

Adapter modules (LM finetuning)

2022-07-22 About

adapter-hub/adapter-transformers: Huggingface Transformers + Adapters

Tags:

2022-07-22 About

Domain adaptation of word embeddings through the exploitation of in-domain corpora and knowledge bases (PhD Thesis 2021)

Tags:

Thèse de Hicham El Boukkouri, univ Paris-Saclay

[Github](https://github.com/helboukkouri/phd-code)

### Goal

Given a target specialized domain, improve the quality of general-domain
word representations using in-domain corpora and/or knowledge bases

### Contributions

#### a method for specializing general-domain embeddings in a [Low-Resource](tag:nlp_low_resource_scenarios) context.

> - train static representations on the task corpus, 
> - resume the
pre-training of general-domain contextual embeddings on the same task corpus,
> - finally, combine both static and contextual representations into
one final model

#### we tackle the issue of using a general-domain vocabulary in a specialized domain

#### Evaluation of  re-training vs training from scratch on specialized corpora using a specialized vocabulary

training from scratch better, but not that much: re-training from a general model
is still appropriate as it is less expensive and leads to comparable, although
slightly lower, performance

#### Regarding subword-based tokenization systems
> we argue that they are inconvenient in practice -> CharacterBERT, a variant
of BERT that uses ELMo’s character-based system instead of WordPieces. More convenient ti use, superior robustness to misspellings

#### Ways to specialize general-domain representations using knowledge bases

a strong baseline using a simple
method relying on graph embeddings and concatenation, using only is_a relation

> both static and contextual embeddings may effectively be
specialized using this simple approach

#### Knowledge Injection Modules (KIM) that
inject the knowledge representations directly within the BERT-like models' architecture

### Notes

> our experiments focused on a single
setting (i.e. the medical domain and the English language)

> meta-embeddings, an approach that consists in combining
different sets of representations for achieving improved performance

2022-03-23 About

[1902.00751] Parameter-Efficient Transfer Learning for NLP

Tags:

2021-04-11 About