Falcon3-prequantized

0. TL;DR
1. Model Details
2. Training Details
3. Usage
4. Evaluation
5. Citation

TL;DR

Model Details

This is a 'hacked' version of tiiuae/Falcon3-10B-Base-1.58bit where model weight scales have been injected into ternary model weights in order to make the model compatible with fine-tuning

Model Description

• Developed by: https://www.tii.ae
• Model type: Causal decoder-only
• Architecture: Pure-transformer - 1.58bit version
• Language(s) (NLP): Mainly English
• License: TII Falcon License 2.0

Training details

The model has been trained following the training strategies from the recent 1-bit LLM HF blogpost and 1-bit LLM paper. For more details about the training protocol of this model, please refer to the Falcon-3 technical report, section Compression.

Usage

Currently to use this model you can either rely on Hugging Face transformers library or BitNet library. You can also play with the model using the falcon-1.58bit playground (only for the 7B instruct version).

🤗 transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tiiuae/Falcon3-10B-Base-1.58bit"

model = AutoModelForCausalLM.from_pretrained(
  model_id,
  torch_dtype=torch.bfloat16,
).to("cuda")

# Perform text generation

BitNet

git clone https://github.com/microsoft/BitNet && cd BitNet
pip install -r requirements.txt
python setup_env.py --hf-repo tiiuae/Falcon3-10B-Base-1.58bit -q i2_s
python run_inference.py -m models/Falcon3-10B-1.58bit/ggml-model-i2_s.gguf -p "You are a helpful assistant" -cnv

Evaluation

We report in the following table our internal pipeline benchmarks:

Note evaluation results are normalized score from v2 leaderboard tasks - reported results of original models in the blogpost are raw scores

Citation

Coming soon ..