StableLM: Stability AI Language Models

“A Stochastic Parrot, flat design, vector art” — Stable Diffusion XL

This repository contains Stability AI's ongoing development of the StableLM series of language models and will be continuously updated with new checkpoints. The following provides an overview of all currently available models. More coming soon.

News

September 29, 2023

• Released StableLM-3B-4E1T model under CC BY-SA-4.0.

August 5, 2023

• Released patched StableLM-Alpha v2 models with 3B and 7B parameters.

April 28, 2023

• Released StableVicuna-13B, our RLHF fine-tune of Vicuna-13B v0, which itself is a fine-tune of LLaMA-13B. Delta weights over the original Llama model is released under (CC BY-NC-SA-4.0).

April 20, 2023

• Released initial set of StableLM-Alpha models, with 3B and 7B parameters. Base models are released under CC BY-SA-4.0.

• Try to chat with our 7B model, StableLM-Tuned-Alpha-7B, on Hugging Face Spaces.

Models

StableLM-3B-4E1T

Technical Report: StableLM-3B-4E1T

StableLM-3B-4E1T is a 3 billion (3B) parameter language model pre-trained under the multi-epoch regime to study the impact of repeated tokens on downstream performance. Given prior success in this area (Tay et al., 2023 and Taylor et al., 2022), we train on 1 trillion (1T) tokens for 4 epochs following the observations of Muennighoff et al. (2023) in "Scaling Data-Constrained Language Models" in which they find "training with up to 4 epochs of repeated data yields negligible changes to loss compared to having unique data." Further inspiration for the token count is taken from "Go smol or go home" (De Vries, 2023), which suggests a 2.96B model trained for 2.85 trillion tokens achieves a similar loss to a Chinchilla compute-optimal 9.87B language model ($k_n = 0.3$).

Size	StableLM-3B-4E1T	Training Tokens	Parameters
3B	checkpoint	4T	2,795,443,200

Model Architecture

The model is a decoder-only transformer similar to the LLaMA (Touvron et al., 2023) architecture with the following modifications:

Parameters	Hidden Size	Layers	Heads	Sequence Length
2,795,443,200	2560	32	32	4096

• Position Embeddings: Rotary Position Embeddings (Su et al., 2021) applied to the first 25% of head embedding dimensions for improved throughput following Black et al. (2022).
• Normalization: LayerNorm (Ba et al., 2016) with learned bias terms as opposed to RMSNorm (Zhang & Sennrich, 2019).
• Tokenizer: GPT-NeoX (Black et al., 2022).

Training Data

The dataset is comprised of a filtered mixture of open-source large-scale datasets available on the HuggingFace Hub: Falcon Refined