README.md

metadata

language: en
tags:
  - pytorch
  - causal-lm
  - language-model
  - text-generation
license: mit
datasets:
  - wikitext
metrics:
  - perplexity
  - loss
library_name: pytorch
pipeline_tag: text-generation
model-index:
  - name: SmolLM-125M
    results:
      - task:
          type: text-generation
          name: Language Modeling
        dataset:
          type: wikitext
          name: WikiText-2
        metrics:
          - type: perplexity
            value: to_be_updated
          - type: loss
            value: to_be_updated

SmolLM-125M: A Lightweight Language Model for Consumer Hardware

This is a 125M parameter language model designed to be trained and run on consumer hardware with limited VRAM (4GB+). The model follows a GPT-style architecture but is optimized for efficiency and memory usage.

Model Details

• Architecture: GPT-style Transformer
• Parameters: 125M
• Context Length: 512 tokens
• Vocabulary: 50,257 tokens (GPT-2 tokenizer)
• Training Data: WikiText-2
• Hardware Requirements: 4GB+ VRAM GPU

Architecture Specifications

• Layers: 12 transformer blocks
• Attention Heads: 12
• Embedding Dimension: 768
• Activation: GELU
• Layer Normalization: Pre-norm

Training Details

• Hardware Used: GTX 1650 (4GB VRAM)
• Training Time: ~4 hours
• Batch Size: 4 (16 with gradient accumulation)
• Learning Rate: 3e-4 with cosine decay
• Weight Decay: 0.1
• Optimizer: AdamW

Memory Optimizations

1. Length-based batch scheduling
2. Gradient accumulation (4 steps)
3. Dynamic batch scheduling
4. Pre-padded sequences

Usage

from transformers import AutoTokenizer
from model import SmallLanguageModel, ModelConfig

# Initialize model
config = ModelConfig(
    vocab_size=50257,
    block_size=512,
    n_layer=12,
    n_head=12,
    n_embd=768,
    dropout=0.1,
    bias=True
)
model = SmallLanguageModel(config)

# Generate text
tokenizer = AutoTokenizer.from_pretrained("gpt2")
input_text = "Once upon a time"
input_ids = tokenizer(input_text, return_tensors="pt").input_ids
output_ids = model.generate(input_ids, max_length=100)
generated_text = tokenizer.decode(output_ids[0])

Limitations

• Limited context window (512 tokens)
• Smaller capacity compared to larger models
• Training data limited to WikiText-2

License

This model is released under the MIT License.