|
|
--- |
|
|
base_model: unsloth/llama-3.2-1b-instruct-unsloth-bnb-4bit |
|
|
library_name: peft |
|
|
license: apache-2.0 |
|
|
datasets: |
|
|
- garage-bAInd/Open-Platypus |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- MATH |
|
|
- LEETCODE |
|
|
- text-generation-inference |
|
|
- SCIENCE |
|
|
--- |
|
|
|
|
|
# Model Card for SicMundus |
|
|
|
|
|
## Model Details |
|
|
|
|
|
### Model Description |
|
|
This model, **Pinnacle**, is a fine-tuned version of `unsloth/Llama-3.2-1B-Instruct` utilizing Parameter Efficient Fine-Tuning (PEFT) with LoRA (Low-Rank Adaptation). It has been trained on the `Open-Platypus` dataset with a structured Alpaca-style prompt format. The primary goal is to enhance instruction-following capabilities while maintaining efficiency through 4-bit quantization. |
|
|
|
|
|
- **Developed by:** Ragul |
|
|
- **Funded by:** Self-funded |
|
|
- **Organization:** Pinnacle Organization |
|
|
- **Shared by:** Ragul |
|
|
- **Model type:** Instruction-tuned Language Model |
|
|
- **Language(s) (NLP):** English |
|
|
- **License:** Apache 2.0 (or specify if different) |
|
|
- **Finetuned from model:** `unsloth/Llama-3.2-1B-Instruct` |
|
|
|
|
|
### Model Sources |
|
|
- **Repository:** [https://huggingface.co/ragul2607/SicMundus] |
|
|
- **Paper:** N/A (or link to relevant research) |
|
|
- **Demo:** [Gradio, HF Spaces, etc.] |
|
|
|
|
|
## Uses |
|
|
|
|
|
### Direct Use |
|
|
- General-purpose instruction-following tasks |
|
|
- Text generation |
|
|
- Code generation assistance |
|
|
- Conversational AI applications |
|
|
|
|
|
### Downstream Use |
|
|
- Further fine-tuning on domain-specific datasets |
|
|
- Deployment in chatbot applications |
|
|
- Text summarization or document completion |
|
|
|
|
|
### Out-of-Scope Use |
|
|
- Not designed for real-time critical applications (e.g., medical or legal advice) |
|
|
- May not be suitable for handling highly sensitive data |
|
|
|
|
|
## Bias, Risks, and Limitations |
|
|
|
|
|
While the model is designed to be a general-purpose assistant, it inherits biases from the pre-trained Llama model and the Open-Platypus dataset. Users should be aware of potential biases in generated responses, particularly regarding sensitive topics. |
|
|
|
|
|
### Recommendations |
|
|
- Use in conjunction with human oversight. |
|
|
- Avoid deploying in high-stakes scenarios without additional testing. |
|
|
|
|
|
## How to Get Started with the Model |
|
|
|
|
|
To use the fine-tuned model, follow these steps: |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
import torch |
|
|
|
|
|
model_path = "path/to/SicMundus" |
|
|
tokenizer = AutoTokenizer.from_pretrained(model_path) |
|
|
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.float16, device_map="auto") |
|
|
|
|
|
def generate_response(prompt): |
|
|
inputs = tokenizer(prompt, return_tensors="pt").to("cuda") |
|
|
output = model.generate(**inputs, max_new_tokens=100) |
|
|
return tokenizer.decode(output[0], skip_special_tokens=True) |
|
|
|
|
|
prompt = "Explain the concept of reinforcement learning." |
|
|
print(generate_response(prompt)) |
|
|
``` |
|
|
|
|
|
## Training Details |
|
|
|
|
|
### Training Data |
|
|
- **Dataset:** `garage-bAInd/Open-Platypus` |
|
|
- **Preprocessing:** The dataset was formatted using Alpaca-style prompts with instruction, input, and output fields. |
|
|
|
|
|
### Training Procedure |
|
|
- **Training Framework:** Hugging Face `transformers` + `trl` (PEFT + LoRA) |
|
|
- **Precision:** Mixed precision (FP16/BF16 based on hardware support) |
|
|
- **Batch size:** 2 per device with gradient accumulation |
|
|
- **Learning rate:** 2e-4 |
|
|
- **Max Steps:** 100 |
|
|
- **Optimizer:** AdamW 8-bit |
|
|
- **LoRA Config:** Applied to key transformer layers (q_proj, k_proj, v_proj, etc.) |
|
|
|
|
|
### Speeds, Sizes, Times |
|
|
- **Checkpoint Size:** ~2GB (LoRA adapters stored separately) |
|
|
- **Fine-tuning Time:** ~1 hour on A100 GPU |
|
|
|
|
|
## Evaluation |
|
|
|
|
|
### Testing Data, Factors & Metrics |
|
|
- **Testing Data:** A subset of Open-Platypus |
|
|
- **Factors:** Performance on general instruction-following tasks |
|
|
- **Metrics:** |
|
|
- Perplexity (PPL) |
|
|
- Response Coherence |
|
|
- Instruction-following accuracy |
|
|
|
|
|
### Results |
|
|
- **Perplexity:** TBD |
|
|
- **Response Quality:** Qualitatively improved over base model on test prompts |
|
|
|
|
|
## Model Examination |
|
|
- **Interpretability:** Standard transformer-based behavior with LoRA fine-tuning. |
|
|
- **Explainability:** Outputs can be analyzed with attention visualization tools. |
|
|
|
|
|
## Environmental Impact |
|
|
- **Hardware Type:** A100 GPU |
|
|
- **Hours used:** ~1 hour |
|
|
- **Cloud Provider:** Local GPU / AWS / Hugging Face Accelerate |
|
|
- **Carbon Emitted:** Estimated using [Machine Learning Impact Calculator](https://mlco2.github.io/impact) |
|
|
|
|
|
## Technical Specifications |
|
|
|
|
|
### Model Architecture and Objective |
|
|
- Transformer-based architecture (Llama-3.2-1B) |
|
|
- Instruction-following optimization with PEFT-LoRA |
|
|
|
|
|
### Compute Infrastructure |
|
|
- **Hardware:** A100 (or specify if different) |
|
|
- **Software:** Python, PyTorch, `transformers`, `unsloth`, `peft` |
|
|
|
|
|
## Citation |
|
|
If using this model, please cite: |
|
|
|
|
|
```bibtex |
|
|
@misc{SicMundus, |
|
|
author = {Ragul}, |
|
|
title = {SicMundus: Fine-Tuned Llama-3.2-1B-Instruct}, |
|
|
year = {2025}, |
|
|
url = {https://huggingface.co/ragul2607/SicMundus} |
|
|
} |
|
|
``` |
|
|
|
|
|
## More Information |
|
|
- **Contact:** [https://github.com/ragultv] |
|
|
- **Further Work:** Integrate with RLHF for better alignment |
|
|
|
|
|
## Model Card Authors |
|
|
- Ragul |
