This repository hosts a layer-pruned version of HuggingFaceTB/SmolLM-135M, reduced from ~ 135M parameters down to ~ 99M parameters (~26% smaller).

⚠️ Note: This model is intended as a starting point for knowledge distillation or fine-tuning, not as a final standalone model.

Model Details

Model Description

Developed by: Independent (based on HuggingFaceTB/SmolLM-135M)
Model type: Decoder-only causal language model (LLaMA-style)
Language(s) (NLP): English (same as base model)
License: Inherits license from HuggingFaceTB/SmolLM-135M
Finetuned from model: HuggingFaceTB/SmolLM-135M

The model was pruned from 30 layers → 20 layers, achieving ~26% parameter reduction while keeping embeddings, the output head, and the final layer intact.

Model Sources

Repository: HuggingFaceTB/SmolLM-135M (base)
This model repo: current repository
Paper [optional]: N/A
Demo [optional]: N/A

Uses

Direct Use

Educational / research purposes for studying pruning effects on transformer models.
Lightweight inference where resources are limited.

Downstream Use

Knowledge Distillation: Using this pruned model as a student model against larger teacher models.
Fine-Tuning: Domain adaptation on specific datasets while benefiting from lower compute requirements.

Out-of-Scope Use

Production deployment without evaluation.
High-stakes applications (medical, legal, safety-critical systems).

Bias, Risks, and Limitations

As with the base model, it may generate biased or toxic text.
Pruning reduces capacity → performance may drop without re-training.
Model has not been benchmarked post-pruning.

Recommendations

Users should:

Perform task-specific fine-tuning or distillation before deployment.
Benchmark against baselines to measure trade-offs in accuracy vs. efficiency.

How to Get Started with the Model

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "your-username/SmolLM-135M-layer-pruned-90M-raw"

tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/SmolLM-135M")
model = AutoModelForCausalLM.from_pretrained(model_id)

inputs = tokenizer("Hello, how are you?", return_tensors="pt")
outputs = model.generate(**inputs, max_length=50)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

abhinavv3/SmolLM-135M-layer-pruned-90M-raw

Model Card for SmolLM-135M Layer-Pruned (~90M params)