| <div align="center"> | |
| # TinyLlama-1.1B | |
| English | [δΈζ](README_zh-CN.md) | |
| </div> | |
| The TinyLlama project aims to **pretrain** a **1.1B Llama model on 3 trillion tokens**. With some proper optimization, we can achieve this within a span of "just" 90 days using 16 A100-40G GPUs ππ. The training has started on 2023-09-01. | |
| <div align="center"> | |
| <img src=".github/TinyLlama_logo.png" width="300"/> | |
| </div> | |
| We adopted exactly the same architecture and tokenizer as Llama 2. This means TinyLlama can be plugged and played in many open-source projects built upon Llama. Besides, TinyLlama is compact with only 1.1B parameters. This compactness allows it to cater to a multitude of applications demanding a restricted computation and memory footprint. | |
| #### Releases Schedule | |
| We will be rolling out intermediate checkpoints following the below schedule. We also include some baseline models for comparison. | |
| | Date | HF Checkpoint | Tokens | Step | HellaSwag Acc_norm | | |
| |------------|-------------------------------------------------|--------|------|---------------------| | |
| | Baseline | [StableLM-Alpha-3B](https://huggingface.co/stabilityai/stablelm-base-alpha-3b)| 800B | -- | 38.31 | | |
| | Baseline | [Pythia-1B-intermediate-step-50k-105b](https://huggingface.co/EleutherAI/pythia-1b/tree/step50000) | 105B | 50k | 42.04 | | |
| | Baseline | [Pythia-1B](https://huggingface.co/EleutherAI/pythia-1b) | 300B | 143k | 47.16 | | |
| | 2023-09-04 | [TinyLlama-1.1B-intermediate-step-50k-105b](https://huggingface.co/PY007/TinyLlama-1.1B-step-50K-105b) | 105B | 50k | 43.50 | | |
| | 2023-09-16 | -- | 500B | -- | -- | | |
| | 2023-10-01 | -- | 1T | -- | -- | | |
| | 2023-10-16 | -- | 1.5T | -- | -- | | |
| | 2023-10-31 | -- | 2T | -- | -- | | |
| | 2023-11-15 | -- | 2.5T | -- | -- | | |
| | 2023-12-01 | -- | 3T | -- | -- | | |
| <!-- | Baseline | [Pythia-1B-intermediate-52b](https://huggingface.co/EleutherAI/pythia-1b/tree/step25000) | 52B | 25k | 38.81 | --> | |
| <!-- | Baseline | [Pythia-1.4B-intermediate-52b](https://huggingface.co/EleutherAI/pythia-1.4b/tree/step25000) | 52B | 25k | 42.49 | --> | |
| <!-- | Baseline | [Pythia-1.4B-intermediate-105b](https://huggingface.co/EleutherAI/pythia-1.4b/tree/step50000) | 105B | 50k | 46.14 | --> | |
| <!-- | 2023-09-04 | [TinyLlama-1.1B-intermediate-52b](https://huggingface.co/PY007/TinyLlama-1.1B-52b) | 52B | 25k | 40.85 | | |
| | 2023-09-04 | [TinyLlama-1.1B-intermediate-84b](https://huggingface.co/PY007/TinyLlama-1.1B-84b) | 84B | 40k | 42.65 | --> | |
| It can be observed that TinyLlama has so far progressed well ππ. | |
| Meanwhile, you can track the live cross entropy loss [here](https://wandb.ai/lance777/lightning_logs/reports/metric-train_loss-23-09-02-15-26-17---Vmlldzo1MjkzNzMw?accessToken=9843chbl7rfi1w03hxttpcnbo9z8t6088pw3ddn4h8teunaq0cy7j8hw9c5i02ve). | |
| ## Training Details | |
| Below are some details of our training setup: | |
| | Setting | Description | | |
| |---------------------------------|----------------------------------------------------------------| | |
| | Parameters | 1.1B | | |
| | Attention Variant | Grouped Query Attention | | |
| | Model Size | Layers: 22, Heads: 32, Query Groups: 4, Embedding Size: 2048, Intermediate Size (Swiglu): 5632| | |
| | Sequence Length | 2048 | | |
| | Batch Size | 2 million tokens (2048 * 1024) | | |
| | Learning Rate | 4e-4 | | |
| | Learning Rate Schedule | Cosine with 2000 warmup steps | | |
| | Training Data | [Slimpajama](https://huggingface.co/datasets/cerebras/slimpajama-627b) & [Starcoderdata](https://huggingface.co/datasets/bigcode/starcoderdata) | | |
| | Data Preprocessing | Excluded GitHub subset of Slimpajama; Sampled all code from Starcoderdata | | |
| | Combined Dataset Size | 1 trillion tokens | | |
| | Total Tokens During Training | 3 trillion (3 epochs/1430k steps) | | |
| | Natural Language to Code Ratio | 7:3 | | |
| | Hardware | 16 A100-40G GPUs | | |