SparkAI-47M-Llama-Instruct
Instruction-tuned checkpoint of SparkAI-47M-Llama (~48M parameter decoder-only transformer), fine-tuned for chat and instruction following.
🌟 Key Highlights & Unique Features
- ⚡ Ultra-Low Memory Footprint (~95.4 MB): Fits in under 100MB of RAM, making it suitable for edge devices, mobile apps, WebGPU, and microcontrollers.
- 🏋️ Data-Saturated Pretraining (10 Billion Tokens): Pretrained on 10B tokens (210 tokens/param) of high-quality FineWeb-Edu + Cosmopedia-v2 text, providing an empirical benchmark on capacity saturation for sub-50M models.
- 🏗️ Modern LLaMA 3 Architecture: Built with Grouped Query Attention (GQA), SwiGLU activations, RoPE positional encodings, RMSNorm pre-normalization, and tied embeddings.
- 💬 Full ChatML SFT Alignment: Fine-tuned with ChatML
<|im_start|>instruction formatting and template support (chat_template.jinja).
📐 Architecture
- Parameters: ~48M (~47.4M non-embedding / tied embeddings)
- Layers: 8
- Hidden Size: 512
- Attention Heads: 8 query heads, 2 key/value heads (Grouped Query Attention - GQA)
- MLP: SwiGLU (Intermediate size: 1408)
- Positional Encoding: RoPE (Rotary Position Embeddings)
- Normalization: RMSNorm (Pre-normalization)
- Embeddings: Tied embeddings, no bias terms
- Vocabulary Size: 49,152 (SmolLM2 tokenizer with Chat Template)
- Sequence Length: 1024
🏋️ Training & Fine-Tuning Details
- Base Checkpoint:
vedantjadhav701/SparkAI-47m-llama-10b-token - Pretraining Data: FineWeb-Edu (
sample-100BT) + Cosmopedia-v2 (85% / 15% mix, 10.00B tokens) - Optimizer: AdamW with Cosine LR schedule + warmup
- Hardware: NVIDIA A100 80GB PCIe
- Final Eval Perplexity: 31.46
📈 Pretraining Progression
| Tokens | Perplexity |
|---|---|
| 630M | 43.49 |
| 3.77B | 31.30 |
| 7.00B | — |
| 10.00B | 31.46 |
Note on Saturation: Perplexity plateaued between 3.77B and 10.00B tokens despite continued training, indicating the model has saturated its representational capacity at this size.
📊 Comparison in Sub-50M Parameter Landscape
| Feature | Typical Sub-50M Models | SparkAI-47M-Llama / Instruct |
|---|---|---|
| Token Budget | ~1B – 2B tokens | 10.00 Billion Tokens (210 tokens/param) |
| Data Quality | Raw web text / C4 | FineWeb-Edu (85%) + Cosmopedia-v2 (15%) |
| Architecture | Basic MHA / GPT-2 style | Modern LLaMA 3 (GQA, SwiGLU, RoPE, RMSNorm) |
| Model Size | ~100MB – 200MB | ~95.4 MB (model.safetensors) |
| SFT Alignment | Rare / None | Instruction-tuned with ChatML (chat_template.jinja) |
| Benchmarking | Few metrics | Empirical capacity saturation documented at 10B tokens |
💻 Usage with Hugging Face transformers
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
repo_id = "vedantjadhav701/SparkAI-47m-llama-instruct"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id)
messages = [
{"role": "user", "content": "What is a computer program?"}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=80, do_sample=True, temperature=0.6)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
🚀 Quickstart & Local Gradio UI
Installation
pip install -r requirements.txt
Run Gradio App
python app.py
Open http://127.0.0.1:7860 in your web browser.