---
library_name: transformers
license: mit
datasets:
- MBZUAI/LaMini-instruction
---
# Saving 77% of the Parameters in Large Language Models Technical Report
This repository contains experiment results for the [Saving 77% of the Parameters in Large Language Models Technical Report (PDF)](https://www.researchgate.net/publication/388835829_SAVING_77_OF_THE_PARAMETERS_IN_LARGE_LANGUAGE_MODELS_TECHNICAL_REPORT).

## Abstract
This technical report demonstrates that large language models (LLMs) can maintain their learning capacity while reducing their non-embedding parameters by up to 77%. We achieve this by adapting a parameter reduction technique originally developed for computer vision, replacing dense layers with an optimized subnetwork that contains grouped pointwise convolutions. Using Microsoft's phi-3-mini-4k-instruct as our baseline, we show that our optimized model (kphi-3) achieves comparable validation loss while using only 15-23% of the original non-embedding parameters. All experiments were conducted on a single NVIDIA L4 GPU within a 3-day timeframe, supporting the democratization of AI research. Our findings suggest that current LLM architectures may be substantially overparameterized, opening possibilities for more efficient model training and deployment.

## Key Findings
- Achieved 77% parameter reduction while maintaining model performance.
- Demonstrated better generalization in optimized models.
- Improved output quality in qualitative testing.

## Implementation Details
- Base Model: [kphi3](https://github.com/joaopauloschuler/less-parameters-llm).
- Training Dataset: [LaMini](https://huggingface.co/datasets/MBZUAI/LaMini-instruction).
- Architecture: Modified transformer decoder with grouped pointwise convolutions.

## Results
The following table shows LaMini training results with the baseline and the optimized versions. From left to right: experiment label, model name, number of transformer decoder layers, intermediate dimensions, number of non-embedding parameters, training loss and validation loss.

| label | model | layers | interm. dims. | non-emb. params. | % | Train Loss | Val. Loss |
|:-----:|:------:|:-------:|:-------------:|:----------------:|:---:|:----------:|:----------:|
| [JP47D54C](https://github.com/joaopauloschuler/less-parameters-llm/tree/main/raw/JP47D54C_Baseline_2T.ipynb) | phi-3 | 2 | 8192 | [227M](https://huggingface.co/schuler/experimental-JP47D54C) | | **1.08** | 1.58 |
| [JP47D55C](https://github.com/joaopauloschuler/less-parameters-llm/tree/main/raw/JP47D55C_kphi3_2T.ipynb) | kphi-3 | 2 | 9216 | [**35M**](https://huggingface.co/schuler/experimental-JP47D55C) | **15%** | 1.26 | 1.60 |
| [JP47D56C](https://github.com/joaopauloschuler/less-parameters-llm/tree/main/raw/JP47D56C_kphi3_3T.ipynb) | kphi-3 | 3 | 9216 | [53M](https://huggingface.co/schuler/experimental-JP47D56C) | 23% | 1.21 | **1.57** |

## Quick Links
- [📄 Full Technical Report (PDF)](https://www.researchgate.net/publication/388835829_SAVING_77_OF_THE_PARAMETERS_IN_LARGE_LANGUAGE_MODELS_TECHNICAL_REPORT)
- [🤗 Model Checkpoints on HuggingFace](https://huggingface.co/schuler/)
- [📊 Raw Experiment Files](https://github.com/joaopauloschuler/less-parameters-llm/tree/main/raw)

## Usage:
```
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig, pipeline
from transformers import LlamaTokenizer
import torch

REPO_NAME = 'schuler/experimental-JP47D56C'

def load_model(local_repo_name):
    tokenizer = LlamaTokenizer.from_pretrained(local_repo_name, trust_remote_code=True)
    generator_conf = GenerationConfig.from_pretrained(local_repo_name)
    model = AutoModelForCausalLM.from_pretrained(local_repo_name, trust_remote_code=True, torch_dtype=torch.bfloat16, attn_implementation="eager")
    # model.to('cuda')
    return tokenizer, generator_conf, model

tokenizer, generator_conf, model = load_model(REPO_NAME)
global_error = ''
try:
  generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
except Exception as e:
  global_error =  f"Failed to load model: {str(e)}"

def PrintTest(str):
  print(generator(str, max_new_tokens=256, do_sample=True, top_p=0.25, repetition_penalty=1.2))

PrintTest(f"<|user|>\nHello\n<|end|>\n<|assistant|>\n")
PrintTest(f"<|user|>Hello\n<|end|><|assistant|>")
PrintTest(f"<|user|>\nWhat is the human body?\n<|end|>\n<|assistant|>\n")
PrintTest(f"<|user|>What is the human body?\n<|end|><|assistant|>")
PrintTest(f"<|user|>What is biology?\n<|end|><|assistant|>")
PrintTest(f"<|user|>Can you comment about democracy?\n<|end|><|assistant|>")
PrintTest(f"<|user|>Can you provide detailed comments about the concept of democracy?\n<|end|><|assistant|>")
PrintTest(f"<|user|>Please give me a detailed description of the python computer language.\n<|end|><|assistant|>")
PrintTest(f"<|user|>If you had a difficult task to complete, how would you complete it?\n<|end|><|assistant|>")
PrintTest(f"<|user|>Before replying to my question, I would like you to provide two candidate solutions and do some reflexion about these solution. Then, you'll pick the best as the final reply. What is best: eating healthy or eating economically?\n<|end|><|assistant|>")
```

## Output Examples
```
```

## Citing this Model 
```
@article{SchulerRojas_2025,
  title={Saving 77% of the Parameters in Large Language Models Technical Report},
  url={https://www.researchgate.net/publication/388835829_SAVING_77_OF_THE_PARAMETERS_IN_LARGE_LANGUAGE_MODELS_TECHNICAL_REPORT},
  author={Schwarz Schuler, Joao Paulo and Rojas Gómez, Alejandra},
  year={2025}}
```