File size: 2,717 Bytes
bb8846f 8c22c3d bb8846f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 |
---
datasets:
- zhiman-ai/alpaca_en_demo
base_model:
- meta-llama/Llama-3.2-3B-Instruct
tags:
- lora
- sft
---
# **LLAMA-3.2-3B-Alpaca_en_LORA_SFT**
This model is a fine-tuned version of [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) using the [alpaca_en_demo](https://huggingface.co/datasets/zhiman-ai/alpaca_en_demo) dataset. The fine-tuning process was conducted by **Sri Santh M** for development purposes.
It achieves the following results on the evaluation set:
- **Loss**: 1.0510
---
## **Model Description**
This model is optimized for tasks involving instruction-following, text generation, and fine-tuned identity-based use cases. It leverages the capabilities of the LLaMA-3.2-3B-Instruct base model with additional refinements made using a lightweight fine-tuning approach via PEFT (Parameter-Efficient Fine-Tuning).
---
### **Intended Uses**
- Instruction-following tasks.
- Conversational AI and question-answering applications.
- Text summarization and content generation.
---
## **Training and Evaluation Data**
The model was fine-tuned using the **alpaca_en_demo** dataset, which is designed for instruction-tuned task completion. This dataset includes diverse English-language tasks for demonstrating instruction-following capabilities.
- **Dataset link**: [alpaca_en_demo](https://huggingface.co/datasets/zhiman-ai/alpaca_en_demo)
Further details on the dataset:
- **Source**: zhiman-ai.
- **Size**: Small-scale, development-focused dataset.
- **Purpose**: Designed to emulate instruction-tuned datasets like Alpaca, with a subset of English-language prompts and responses.
---
## **Training Procedure**
### **Hyperparameters**
- **Learning rate**: 0.0001
- **Train batch size**: 1
- **Eval batch size**: 1
- **Gradient accumulation steps**: 8
- **Total effective batch size**: 8
- **Optimizer**: AdamW (torch)
- **Betas**: (0.9, 0.999)
- **Epsilon**: 1e-08
- **Learning rate scheduler**: Cosine schedule with 10% warmup.
- **Number of epochs**: 3.0
### **Frameworks and Libraries**
- **PEFT**: 0.12.0
- **Transformers**: 4.46.1
- **PyTorch**: 2.4.0
- **Datasets**: 3.1.0
- **Tokenizers**: 0.20.3
---
## **Training Results**
- **Loss**: 1.0510
- Evaluation results are limited to the dataset scope. Broader testing is recommended for downstream applications.
---
## **Additional Information**
- **Author**: Sri Santh M
- **Purpose**: Fine-tuned for development and experimentation purposes using the LLaMA-3.2-3B-Instruct model.
This model serves as an experimental proof-of-concept for lightweight fine-tuning using PEFT and can be adapted further based on specific tasks or use cases. |