Multiple GPTQ parameter permutations are provided; see Provided Files below for details of the options provided, their parameters, and the software used to create them.
These files were quantised using hardware kindly provided by Massed Compute.
This may not be a complete list; if you know of others, please let me know!
Provided files, and GPTQ parameters
Multiple quantisation parameters are provided, to allow you to choose the best one for your hardware and requirements.
Each separate quant is in a different branch. See below for instructions on fetching from different branches.
Most GPTQ files are made with AutoGPTQ. Mistral models are currently made with Transformers.
Explanation of GPTQ parameters
Bits: The bit size of the quantised model.
GS: GPTQ group size. Higher numbers use less VRAM, but have lower quantisation accuracy. "None" is the lowest possible value.
Act Order: True or False. Also known as desc_act. True results in better quantisation accuracy. Some GPTQ clients have had issues with models that use Act Order plus Group Size, but this is generally resolved now.
Damp %: A GPTQ parameter that affects how samples are processed for quantisation. 0.01 is default, but 0.1 results in slightly better accuracy.
GPTQ dataset: The calibration dataset used during quantisation. Using a dataset more appropriate to the model's training can improve quantisation accuracy. Note that the GPTQ calibration dataset is not the same as the dataset used to train the model - please refer to the original model repo for details of the training dataset(s).
Sequence Length: The length of the dataset sequences used for quantisation. Ideally this is the same as the model sequence length. For some very long sequence models (16+K), a lower sequence length may have to be used. Note that a lower sequence length does not limit the sequence length of the quantised model. It only impacts the quantisation accuracy on longer inference sequences.
ExLlama Compatibility: Whether this file can be loaded with ExLlama, which currently only supports Llama and Mistral models in 4-bit.
If you remove the --local-dir-use-symlinks False parameter, the files will instead be stored in the central Hugging Face cache directory (default location on Linux is: ~/.cache/huggingface), and symlinks will be added to the specified --local-dir, pointing to their real location in the cache. This allows for interrupted downloads to be resumed, and allows you to quickly clone the repo to multiple places on disk without triggering a download again. The downside, and the reason why I don't list that as the default option, is that the files are then hidden away in a cache folder and it's harder to know where your disk space is being used, and to clear it up if/when you want to remove a download model.
The cache location can be changed with the HF_HOME environment variable, and/or the --cache-dir parameter to huggingface-cli.
Note that using Git with HF repos is strongly discouraged. It will be much slower than using huggingface-hub, and will use twice as much disk space as it has to store the model files twice (it stores every byte both in the intended target folder, and again in the .git folder as a blob.)
Python code example: inference from this GPTQ model
Install the necessary packages
Requires: Transformers 4.33.0 or later, Optimum 1.12.0 or later, and AutoGPTQ 0.4.2 or later.
shell
1pip3 install --upgrade transformers optimum
2# If using PyTorch 2.1 + CUDA 12.x:3pip3 install --upgrade auto-gptq
4# or, if using PyTorch 2.1 + CUDA 11.x:5pip3 install --upgrade auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
If you are using PyTorch 2.0, you will need to install AutoGPTQ from source. Likewise if you have problems with the pre-built wheels, you should try building from source:
I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.
If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.
Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.
Patreon special mentions: Brandon Frisco, LangChain4j, Spiking Neurons AB, transmissions 11, Joseph William Delisle, Nitin Borwankar, Willem Michiel, Michael Dempsey, vamX, Jeffrey Morgan, zynix, jjj, Omer Bin Jawed, Sean Connelly, jinyuan sun, Jeromy Smith, Shadi, Pawan Osman, Chadd, Elijah Stavena, Illia Dulskyi, Sebastain Graf, Stephen Murray, terasurfer, Edmond Seymore, Celu Ramasamy, Mandus, Alex, biorpg, Ajan Kanaga, Clay Pascal, Raven Klaugh, 阿明, K, ya boyyy, usrbinkat, Alicia Loh, John Villwock, ReadyPlayerEmma, Chris Smitley, Cap'n Zoog, fincy, GodLy, S_X, sidney chen, Cory Kujawski, OG, Mano Prime, AzureBlack, Pieter, Kalila, Spencer Kim, Tom X Nguyen, Stanislav Ovsiannikov, Michael Levine, Andrey, Trailburnt, Vadim, Enrico Ros, Talal Aujan, Brandon Phillips, Jack West, Eugene Pentland, Michael Davis, Will Dee, webtim, Jonathan Leane, Alps Aficionado, Rooh Singh, Tiffany J. Kim, theTransient, Luke @flexchar, Elle, Caitlyn Gatomon, Ari Malik, subjectnull, Johann-Peter Hartmann, Trenton Dambrowitz, Imad Khwaja, Asp the Wyvern, Emad Mostaque, Rainer Wilmers, Alexandros Triantafyllidis, Nicholas, Pedro Madruga, SuperWojo, Harry Royden McLaughlin, James Bentley, Olakabola, David Ziegler, Ai Maven, Jeff Scroggin, Nikolai Manek, Deo Leter, Matthew Berman, Fen Risland, Ken Nordquist, Manuel Alberto Morcote, Luke Pendergrass, TL, Fred von Graf, Randy H, Dan Guido, NimbleBox.ai, Vitor Caleffi, Gabriel Tamborski, knownsqashed, Lone Striker, Erik Bjäreholt, John Detwiler, Leonard Tan, Iucharbius
Thank you to all my generous patrons and donaters!
And thank you again to a16z for their generous grant.
Original model card: OpenLLM France's Claire 7B 0.1
Claire-7B-0.1
Claire-7B-0.1 is a 7B parameter causal decoder-only model built by LINAGORA and OpenLLM-Franceadapted from Falcon-7b on French conversational data.
Claire-7B-0.1 is a pretrained language model designed to be attuned to the dynamics of linguistic interactions in dialogue. Without further training, its expected use is to generate continuations of dialogues. Its main purpose is to serve as a base model for fine-tuning on dialogue generation (e.g., chat) and dialogue understanding (e.g., meeting summarization) tasks. Please note that due to its training, the model is prone to generate dialogues with disfluencies and other constructions common to spoken language.
Typical usage
python
1import transformers
2import torch
34model_name ="OpenLLM-France/Claire-7B-0.1"56tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)7model = transformers.AutoModelForCausalLM.from_pretrained(model_name,8 device_map="auto",9 torch_dtype=torch.bfloat16,10 load_in_4bit=True# For efficient inference, if supported by the GPU card11)1213pipeline = transformers.pipeline("text-generation", model=model, tokenizer=tokenizer)14generation_kwargs =dict(15 num_return_sequences=1,# Number of variants to generate.16 return_full_text=False,# Do not include the prompt in the generated text.17 max_new_tokens=200,# Maximum length for the output text.18 do_sample=True, top_k=10, temperature=1.0,# Sampling parameters.19 pad_token_id=tokenizer.eos_token_id,# Just to avoid a harmless warning.20)2122prompt ="""\
23- Bonjour Dominique, qu'allez-vous nous cuisiner aujourd'hui ?
24- Bonjour Camille,\
25"""26completions = pipeline(prompt,**generation_kwargs)27for completion in completions:28print(prompt +" […]"+ completion['generated_text'])
This will print something like:
- Bonjour Dominique, qu'allez-vous nous cuisiner aujourd'hui ?
- Bonjour Camille, […] je vous prépare un plat de saison, une daube provençale.
- Ah je ne connais pas cette recette.
- C'est très facile à préparer, vous n'avez qu'à mettre de l'eau dans une marmite, y mettre de l'oignon émincé, des carottes coupées en petits morceaux, et vous allez mettre votre viande de bœuf coupé en petits morceaux également.
- Je n'ai jamais cuisiné de viande de bœuf, mais c'est vrai que ça a l'air bien facile.
- Vous n'avez plus qu'à laisser mijoter, et ensuite il sera temps de servir les clients.
- Très bien.
You will need at least 6GB of VRAM to run inference using 4bit quantization (16GB of VRAM without 4bit quantization).
If you have trouble running this code, make sure you have recent versions of torch, transformers and accelerate (see requirements.txt).
Typical prompts
Claire-7B-0.1 was trained on diarized French conversations. During training, the dialogues were normalized in several formats. The possible formats for expected prompts are as follows:
A monologue can be specified as a single line prompt (though keep in mind that Claire might still return a dialogue because of its training):
prompt = "Mesdames et messieurs les députés, chers collègues, bonsoir. Vous l'aurez peut-être remarqué, je cite rarement"
A dialogue between two speakers can be specified with one line per speech turn starting with a dash:
python
1prompt ="""\
2- Bonjour Dominique, qu'allez-vous nous cuisiner aujourd'hui ?
3- Bonjour Camille,\
4"""
A dialogue or multilogue (with two or more speakers) can be specified with lines that start with [Intervenant X:] where X is a number:
python
1prompt ="""\
2[Intervenant 1:] Bonjour Dominique, qu'allez-vous nous cuisiner aujourd'hui ?
3[Intervenant 2:] Bonjour Camille,\
4"""
A dialogue or multilogue with named speakers can be specified with lines that start with [SpeakerName:]
where SpeakerName can be a first name, a first and a last name, a nickname, a title…
python
1prompt ="""\
2[Mme Camille Durand:] Bonjour Dominique, qu'allez-vous nous cuisiner aujourd'hui ?
3[Mr. Dominique Petit:] Bonjour Camille,\
4"""
Training Details
Training Data
Claire-7B-0.1 was tuned from Falcon-7b on the following data distribution:
Training data was augmented with the following techniques:
varying the format used to indicate speech turns (dashes or [XXX:])
substituting [Intervenant X:] for [SpeakerName:] or vice versa, where [SpeakerName:] might be a real name or a randomly generated name
removing punctuation marks and/or casing (to prepare the model for transcripts produced by some Automatic Speech Recognition systems)
Long conversations were truncated at a maximum of 2048 tokens. Where possible, they were split between speaker turns.
While the model has been trained and evaluated only on French dialogues, it may be able to generate conversations in other languages from the original Falcon-7b training data.
Training Procedure
Claire-7B-0.1 is a causal decoder-only model trained on a causal language modeling task (i.e., predict the next token).
See Falcon-7b for more details.
Claire-7B-0.1 was trained on 1 A100 80GB GPU for about 50 GPU hours.
Hyperparameters were the following:
Hyperparameter
Value
Precision
bfloat16
Optimizer
AdamW
Learning rate
1e-4
Weight decay
1e-2
Batch size
132
LoRA rank
16
LoRA alpha
32
Dropout
0.05
gradient clipping
1
Evaluation
To evaluate Claire-7B-0.1’s ability to generate natural sounding, French conversations, we compared its responses to a variety of prompts with those of three other models:
Claire-Mistral-7B-0.1 (a version of Mistral-7B-v0.1 adapted in the same fashion as Claire-7B-0.1)
We tested an even mixture of monologue and dialogue-style prompts.
Each of the four generated responses was evaluated along three dimensions:
Interaction, Fluency and Relevance.
Evaluators were also asked to rank the four responses by preference.
Our results confirm that continual pre-training of Falcon-7b and Mistral-7B-v0.1 leads to improvement (relative to the base models) along all three evaluation dimensions and that Claire-7B-0.1 outperforms the adapted Mistral counterpart in the Fluency and Relevance categories
(and in the Interaction category if we focus on dialogue-style prompts).
Ranking results also reveal a clear subjective preference for Claire-7B-0.1,
as shown in the following table:
Please note that the model can generate disfluencies and humorous responses as a result of its training on spoken and theatrical text.
More evaluation details will be provided in a separate publication.
License
Given that some of the corpora used for training are only available under CC-BY-NC-SA licenses,
Claire-7B-0.1 is made available under the CC-BY-NC-SA 4.0 license.
This work was performed using HPC resources from GENCI–IDRIS (Grant 2023-AD011014561).
Claire-7B-0.1 was created by members of LINAGORA (in alphabetical order): Ismaïl Harrando, Julie Hunter, Jean-Pierre Lorré, Jérôme Louradour, Michel-Marie Maudet, Virgile Rennard, Guokan Shang.
Special thanks to partners from the OpenLLM-France community, especially Christophe Cerisara (LORIA), Pierre-Carl Langlais and Anastasia Stasenko (OpSci), and Pierre Colombo, for valuable advice.
