This repository is still under development and may still contain various bugs.
This is the official repository for the paper Inheritune: Training Smaller Yet More Attentive Language Models.
Large Language Models (LLMs) have achieved remarkable performance across various natural language processing tasks, primarily due to the transformer architecture and its self-attention mechanism. However, we observe that in standard decoder-style LLMs, attention matrices degenerate to single-column for deeper layers. Layers in this state are unable to learn anything meaningful and mostly redundant; we refer to these as lazy layers. The goal of this paper is to train smaller models by eliminating this structural inefficiency without compromising performance.
Motivated by this observation, we propose Inheritune, a simple yet effective training recipe for developing smaller, high-performing language models. Smaller models trained with Inheritune inherit early transformer layers from a larger pre-trained model, then retrain and progressively expand until they match or exceed the performance of the larger model. We demonstrate that Inheritune enables the training of various sizes of GPT-2 models on datasets like OpenWebText-9B and FineWeb_Edu. Models trained with \method{}, despite having significantly fewer layers, match or even surpass the performance of their larger counterparts. For instance, our 16-layer GPT-2 medium variant achieves comparable performance to the standard 24-layer GPT-2 medium model.
An analysis of a 36-layer GPT-2 large shows the max rank of the attention matrices across all layers.
A closer look at the same GPT-2 large model reveals that the dominant mass proportion of several attention matrices are concentrated in a single column, particularly in deeper layers.
Model derived using Inheritune converges faster and matches the final validation loss of the full-sized model trained from scratch, despite being smaller. Training GPT-2 xlarge vanilla models from scratch and our variants with OpenWebText-9B for 100K steps.
Models derived using Inheritune converge faster and match the final validation loss of the full-sized model despite using fewer layers.
| Models | Recipe | Layers | ARCE (acc) | PIQA (acc) | SciQ (acc) | Hellaswag (acc norm) | Lambada (acc) | Average |
|---|---|---|---|---|---|---|---|---|
| GPT-2 Medium | ||||||||
| rand init | 24 | 51.05 | 61.81 | 74.8 | 30.79 | 20.28 | 47.74 | |
| rand init | 16 | 49.92 | 61.92 | 73.3 | 29.56 | 19.54 | 46.84 | |
| Ours | 16 | 51.26 | 61.81 | 73.8 | 30.55 | 23 | 48.08 | |
| GPT-2 Large† | ||||||||
| rand init | 32 | 52.48 | 64.58 | 75.3 | 32.65 | 22.2 | 49.44 | |
| rand init | 16 | 50.34 | 63.11 | 75 | 30.86 | 21.56 | 48.17 | |
| Ours | 16 | 52.9 | 63.55 | 76.1 | 32.14 | 24.06 | 49.75 |
Models trained with Inheritune outperform both their larger and same-size counterparts trained from scratch on average zero-shot downstream performance.
For evaluation, we use accuracy (acc) and normalized accuracy (acc norm) metrics following the Open LLM leaderboard. All models are trained with FineWeb_edu.
Performance of our 1.5B base LM derived using 1B data with Inheritune on an average of 9 different datasets (left) and MMLU benchmark (right) that evaluates commonsense, truthfulness, natural language inference and language understanding. We compare our model's performance with reference model-OpenLLamA-3B (2x size), other small base LMs of size 1B-2B parameters such as MPT-1.3B, OPT-1.3B, Pythia-1.4B (pre-trained from scratch) and ShearLLaMA-1.5B (pruned and continually trained using existing large base LM).
Below is the comparison of our target model with reference models and other baseline models of similar size when pre-trained from scratch and pre-trained with inherited weights and pruning. Our model, although trained with fewer tokens, achieves comparable performance. We have highlighted all scores where our model achieves at least 90% of the score compared to its reference language model or outperforms at least two of the baseline models. All tasks are evaluated using 0-shot except MMLU, which is 5-shot. The models marked with n/a are trained from scratch.
| Model | Commonsense Reasoning | |||||
|---|---|---|---|---|---|---|
| Name (# train tokens) | Reference | Winograd | PIQA | Boolq | WinoGrande | Logiqa |
| OpenLLaMA-3B (1T) | n/a | 63.46 | 74.97 | 67.18 | 62.27 | 28.4 |
| OPT-1.3B (300B) | n/a | 38.46 | 71.82 | 57.83 | 59.51 | 27.04 |
| Pythia-1.4B (300B) | n/a | 36.54 | 70.89 | 63.12 | 56.99 | 27.65 |
| MPT-1.3B (200B) | n/a | 63.46 | 71.44 | 50.89 | 58.09 | 28.26 |
| Sheared LLaMA-1.3B (50B) | LLaMA2-7B | 36.54 | 73.45 | 62.02 | 58.17 | 27.34 |
| Ours-1.5B (1B) | OpenLLaMA-3B | 50.96 | 56.47 | 61.68 | 51.69 | 25.19 |
| Model | Lang. Understanding & Inference | Factuality | ||||
|---|---|---|---|---|---|---|
| Name (# train tokens) | Reference | MMLU(5) | WNLI | QNLI | MNLI | TruthfulQA |
| OpenLLaMA-3B (1T) | n/a | 27.21 | 50.7 | 51.3 | 37.3 | 35 |
| OPT-1.3B (300B) | n/a | 24.96 | 42.25 | 51.29 | 35.82 | 38.67 |
| Pythia-1.4B (300B) | n/a | 25.56 | 53.52 | 49.48 | 32.76 | 38.66 |
| MPT-1.3B (200B) | n/a | 25.82 | 40.85 | 50.52 | 35.93 | 38.68 |
| Sheared LLaMA-1.3B (50B) | LLaMA2-7B | 25.71 | 49.3 | 50.98 | 37.94 | 37.14 |
| Ours-1.5B (1B) | OpenLLaMA-3B | 25.67 | 43.66 | 49.41 | 34.42 | 48.61 |
[2024-04-22] We've released the first version of codebase for Inheritune in low data regime and also full data regime.
[2024-04-22] We've added the discussions option at the top for community feedback and discussions. Feel free to suggest new experiments and post your results.
If you find this work helpful, please consider citing us:
@inproceedings{Sanyal2024pretraining,
title = {Inheritune: Training Smaller Yet More Attentive Language Models},
author = {Sunny Sanyal, Ravid Shwartz-Ziv, Alexandros G. Dimakis and Sujay Sanghavi},
year = {2024},
url={https://arxiv.org/abs/2404.08634}
}
The training code for small language model 1B-2B is mainly adapted from litgpt. The code for GPT2 experiments are mainly adapted from Sophia and nanoGPT.
The llama image is created using DALLE.