|
|
--- |
|
|
language: |
|
|
- en |
|
|
license: mit |
|
|
tags: |
|
|
- llm |
|
|
- decoder-only |
|
|
- transformer |
|
|
- from-scratch |
|
|
- research |
|
|
- educational |
|
|
- 80m |
|
|
- pytorch |
|
|
- pretraining |
|
|
- custom-architecture |
|
|
pipeline_tag: text-generation |
|
|
inference: |
|
|
parameters: |
|
|
temperature: 0.7 |
|
|
top_p: 0.95 |
|
|
--- |
|
|
|
|
|
# π§ Mini-LLM β 80M Parameter Transformer (Pretrained From Scratch) |
|
|
|
|
|
[]() |
|
|
[]() |
|
|
|
|
|
**Mini-LLM** is an 80M parameter decoder-only transformer trained **fully from scratch** using a custom tokenizer, custom architecture, and custom training loop. |
|
|
It is designed as an educational + research-friendly minimal LLM that demonstrates how modern LLM components are built end-to-end. |
|
|
|
|
|
--- |
|
|
|
|
|
## β¨ Key Features |
|
|
|
|
|
- **80M parameters** β compact but fully functional LLM |
|
|
- **Trained from scratch** (no borrowed checkpoints) |
|
|
- Custom **Byte-Level BPE tokenizer (32k vocab)** |
|
|
- Modern architecture components: |
|
|
- RoPE (Rotary Position Embeddings) |
|
|
- RMSNorm |
|
|
- SwiGLU FeedForward layer |
|
|
- FlashAttention (via PyTorch SDPA) |
|
|
- GQA-ready Attention implementation |
|
|
- **2B tokens** mixed corpus (FineWeb + WikiText + Wikipedia) |
|
|
- Training logs, checkpoints, plots all included for transparency |
|
|
- Released under a permissive license for research & learning |
|
|
|
|
|
--- |
|
|
|
|
|
## π Model Architecture |
|
|
|
|
|
| Component | Value | |
|
|
|----------|-------| |
|
|
| Type | Decoder-only transformer | |
|
|
| Parameters | ~80M | |
|
|
| Layers | 16 | |
|
|
| Embedding dim | 384 | |
|
|
| Attention heads | 6 | |
|
|
| KV Heads | 6 | |
|
|
| MLP Hidden Dim | 1536 (SwiGLU) | |
|
|
| Max sequence length | 2048 | |
|
|
| Norm | RMSNorm | |
|
|
| Positional Encoding | RoPE | |
|
|
| Tokenizer | SentencePiece BPE (32k vocab, byte fallback) | |
|
|
|
|
|
--- |
|
|
|
|
|
## π¦ Files in This Repo |
|
|
|
|
|
- `checkpoints/` β Pretrained model state_dict + optimizer |
|
|
- `safetensors/` β Final consolidated .safetensors file |
|
|
- `logs/` β Training logs in JSONL |
|
|
- `plots/` β Train/val loss curves |
|
|
- `tokenizer.json` β HF-compatible tokenizer |
|
|
- `spm.model` β SentencePiece model |
|
|
|
|
|
--- |
|
|
|
|
|
## π§ͺ Quick Usage (HF Transformers) |
|
|
|
|
|
```python |
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained("Ashx098/Mini-LLM", trust_remote_code=True) |
|
|
tok = AutoTokenizer.from_pretrained("Ashx098/Mini-LLM") |
|
|
|
|
|
prompt = "Hello, how are you?" |
|
|
inputs = tok(prompt, return_tensors="pt") |
|
|
|
|
|
outputs = model.generate(**inputs, max_new_tokens=50) |
|
|
print(tok.decode(outputs[0], skip_special_tokens=True)) |
|
|
``` |
|
|
|
|
|
## π Training Details |
|
|
|
|
|
### Optimizer |
|
|
- **AdamW** (Ξ²1=0.9, Ξ²2=0.95, weight decay=0.1) |
|
|
- **Learning rate**: 6e-4 (cosine annealing + warmup) |
|
|
|
|
|
### Batch β¨ Sequence |
|
|
- **Global batch size** = 32 |
|
|
- **Sequence length** = 2048 |
|
|
- **Gradient accumulation** = 8 |
|
|
|
|
|
### Hardware |
|
|
- Trained on 1Γ NVIDIA A100 80GB |
|
|
|
|
|
## π Training Curve |
|
|
<p align="center"> <img src="https://huggingface.co/Ashx098/Mini-LLM/resolve/main/phase-1-pretraining/plots/loss_curve.png" width="500"> </p> |
|
|
|
|
|
Final loss reached: ~3.25 |
|
|
|
|
|
## π¬ Example Outputs |
|
|
|
|
|
**Prompt**: "Hello, how are you" |
|
|
**Output**: "Hello, how are you?" |
|
|
|
|
|
**Prompt**: "Python is a programming language that" |
|
|
**Output**: "Python is a programming language that allows the history..." |
|
|
|
|
|
## β οΈ Limitations |
|
|
- Small model β limited reasoning, hallucination likely |
|
|
- Not instruction-tuned |
|
|
- Not suitable for production usage |
|
|
- Best viewed as a learning + research artifact |
|
|
|
|
|
## π License |
|
|
MIT License β free for research, modification, and further training. |
|
|
|
|
|
## π Credits |
|
|
Developed by **Avinash Mynampati** |
|
|
Built from scratch using PyTorch + custom training pipeline. |
|
|
|
|
|
### Want to fine-tune or extend it? |
|
|
You can: |
|
|
- Train further with your own dataset |
|
|
- Add LoRA adapters |
|
|
- Use it to learn attention, RoPE, SwiGLU, etc. |
|
|
- Build a tiny instruction-tuned version (coming soon!) |
|
|
|
|
|
## π¬ Contact |
|
|
For questions or collaborations: |
|
|
- **GitHub**: [Ashx098](https://github.com/Ashx098) |
|
|
- **LinkedIn**: [Avinash Mynampati](https://linkedin.com/in/avinash-mynampati) |
|
|
|