|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
pipeline_tag: text-generation |
|
|
library_name: transformers |
|
|
tags: |
|
|
- i3-architecture |
|
|
- custom_code |
|
|
--- |
|
|
|
|
|
# i3-tiny |
|
|
|
|
|
**i3-tiny** is a compact, efficient character-level language model designed for experimentation and exploration in text generation. Despite its small size, it can generate sequences that are quirky, unpredictable, and full of "human-like" character-level errors. |
|
|
|
|
|
--- |
|
|
|
|
|
## Model Overview |
|
|
|
|
|
i3-tiny is trained to predict the next character in a sequence, making it ideal for **character-level language modeling**, **creative text generation**, and **research on lightweight, efficient models**. Its small footprint allows rapid experimentation, even on modest hardware, and it provides a playground for studying how models learn patterns in sequences of characters. |
|
|
|
|
|
The model is **intentionally experimental** — it's not aligned, fact-checked, or polished. Outputs may be coherent, partially readable, or amusingly garbled. |
|
|
|
|
|
--- |
|
|
|
|
|
## Architecture: i3 |
|
|
|
|
|
The **i3 architecture** (pronounced "i-three") is a novel hybrid design optimized for extreme efficiency on resource-constrained hardware. The name reflects its design goal: to enable language model training on modest consumer CPUs, including Intel Core i3 processors. |
|
|
|
|
|
### Key Design Principles |
|
|
|
|
|
i3 combines multiple efficiency techniques to achieve sub-1GB memory usage during training: |
|
|
|
|
|
- **Hybrid sequence modeling**: Blends different approaches to long-range dependency capture, balancing expressiveness with computational efficiency |
|
|
- **Low-rank parameterization**: Strategic use of matrix factorization reduces memory footprint while maintaining model capacity |
|
|
- **Factorized attention mechanisms**: Efficient approximations that preserve attention's ability to model relationships without quadratic memory costs |
|
|
- **Linear-time operations**: Emphasis on operations that scale linearly with sequence length rather than quadratically |
|
|
|
|
|
### Efficiency Characteristics |
|
|
|
|
|
- **Training memory**: < 1 GB RAM total (including model, gradients, and optimizer state) |
|
|
- **Model size**: 711,106 parameters (~2.7 MB in FP32) |
|
|
- **Training speed**: ~450 ms per iteration on modest CPU hardware |
|
|
- **Sequence processing**: Linear complexity enables longer context windows on limited hardware |
|
|
|
|
|
The architecture is designed from the ground up for CPU-friendly training, making it accessible for experimentation and research without requiring specialized hardware. |
|
|
|
|
|
--- |
|
|
|
|
|
## Training Details |
|
|
|
|
|
* **Dataset:** ~45,830 characters (a curated text corpus repeated for exposure) |
|
|
* **Vocabulary:** 34 characters (all lowercased) |
|
|
* **Sequence length:** 128 |
|
|
* **Training iterations:** 2,000 |
|
|
* **Batch size:** 2 |
|
|
* **Optimizer:** AdamW, learning rate 3e-4 |
|
|
* **Model parameters:** 711,106 |
|
|
* **Hardware:** Trained on free-tier CPU compute (Kaggle) |
|
|
* **Performance notes:** Each iteration takes roughly 400–500 ms; 100 iterations take ~45 s on average. Loss steadily decreased from 3.53 to 2.15 over training. |
|
|
|
|
|
### Training Analysis |
|
|
|
|
|
The charts below illustrate the model's performance over the 2,000 training iterations. |
|
|
|
|
|
The **Training Loss Over Iterations** plot shows a clear learning trend, with the 50-iteration moving average (red line) confirming a steady decrease in Cross-Entropy loss from $\sim3.5$ to $\sim2.1$. The **Training Time Performance** plot shows a consistent block time per 100 iterations, resulting in a nearly linear increase in cumulative training time, demonstrating stable and predictable training execution. |
|
|
|
|
|
 |
|
|
|
|
|
**Example generation (iteration 1200):** |
|
|
|
|
|
``` |
|
|
Prompt: "The quick" |
|
|
Generated: the quick efehn. dethe cans the fice the fpeens antary of eathetint, an thadat hitimes the and cow thig, and |
|
|
``` |
|
|
|
|
|
These outputs capture the **chaotic creativity** of a character-level model: a mixture of readable words, invented forms, and surprising sequences. |
|
|
|
|
|
--- |
|
|
|
|
|
## Use Cases |
|
|
|
|
|
- **Educational research**: Study how tiny models learn language patterns |
|
|
- **Creative text generation**: Experiment with character-level generation |
|
|
- **Efficiency benchmarking**: Test memory-constrained training scenarios |
|
|
- **Architecture research**: Explore novel approaches to efficient language modeling |
|
|
|
|
|
--- |
|
|
|
|
|
## Limitations |
|
|
|
|
|
- Character-level modeling only (no tokenization) |
|
|
- Small vocabulary (34 characters) |
|
|
- Limited training data and iterations |
|
|
- Not suitable for production use or factual tasks |
|
|
- Outputs are experimental and unfiltered |
|
|
|
|
|
--- |
|
|
|
|
|
## Citation |
|
|
|
|
|
If you use this model or the i3 architecture in your research, please cite: |
|
|
|
|
|
```bibtex |
|
|
@misc{i3tiny2024, |
|
|
author = {FlameF0X}, |
|
|
title = {i3-tiny: Ultra-Efficient Character-Level Language Model}, |
|
|
year = {2024}, |
|
|
publisher = {HuggingFace}, |
|
|
howpublished = {\url{https://huggingface.co/FlameF0X/i3-tiny}} |
|
|
} |
|
|
``` |
