AmanPriyanshu/Dynamic-Topic-Modeling-Llama-3.2-1B-bnb-4bit
This repository contains code and documentation for the Llama 3.2 1B variant of our dynamic topic modeling series, based on the RedPajama dataset subset we created for dynamic topic-modeling. Link to dataset: AmanPriyanshu/Dynamic-Topic-RedPajama-Data-1T-100k-SubSample-max-1k-tokens
Model Comparisons
| Model | Parameters | Loss |
|---|---|---|
| Dynamic-Topic-Modeling-Llama-3.2-1B-bnb-4bit | 1B | 2.3959 |
Model Overview
- Base Model: Unsloth/Llama-3.2-1B-bnb-4bit
- Fine-tuned Version: AmanPriyanshu/Dynamic-Topic-Modeling-Llama-3.2-1B-bnb-4bit
- Task: Hierarchical Topic Generation
- Training Data: 100k samples from RedPajama-1T
Dataset Details
The model was trained on a carefully curated subset of the RedPajama-1T dataset:
- 100,000 documents
- Maximum 1,024 tokens per document
- Three-level hierarchical topic annotations
- Original sources include: CommonCrawl, C4, GitHub, Books, ArXiv, Wikipedia, StackExchange
Model Architecture & Training
Key Configuration:
- Sequence Length: 2048 tokens
- LoRA Parameters:
- Rank: 16
- Alpha: 16
- Target Modules: q_proj, k_proj, v_proj, up_proj, down_proj, o_proj, gate_proj
- RSLoRA enabled
Training Parameters:
- Batch Size: 4
- Gradient Accumulation Steps: 2
- Learning Rate: 3e-4
- Epochs: 1
- Optimizer: AdamW 8-bit
- Weight Decay: 0.01
- Warmup Steps: 10
Usage
Check out the Colab Notebook: https://colab.research.google.com/drive/173dKPjMmFZcbMoyHHdG5pOWH5qRyhcyW?usp=sharing
Installation
pip install unsloth transformers torch
Inference Code
from unsloth import FastLanguageModel
import torch
from transformers import TextStreamer
class LlamaInference:
def __init__(self, model_path: str, device: str = "cuda"):
self.device = device
self.model, self.tokenizer = FastLanguageModel.from_pretrained(
model_name=model_path,
max_seq_length=2048,
load_in_4bit=True,
dtype=None,
)
self.model = FastLanguageModel.for_inference(self.model)
self.model.eval()
def generate_response(
self,
prompt: str,
max_new_tokens: int = 512,
temperature: float = 0.7,
top_p: float = 0.9,
stream: bool = True
) -> str:
messages = [{"from": "human", "value": prompt}]
inputs = self.tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt"
).to(self.device)
streamer = TextStreamer(self.tokenizer) if stream else None
with torch.no_grad():
outputs = self.model.generate(
input_ids=inputs,
max_new_tokens=max_new_tokens,
temperature=temperature,
top_p=top_p,
streamer=streamer,
use_cache=True,
pad_token_id=self.tokenizer.pad_token_id,
eos_token_id=self.tokenizer.eos_token_id,
)
if not stream:
response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
response = response.split("assistant\n")[-1].strip()
return response
Example Usage
model_path = "AmanPriyanshu/Dynamic-Topic-Modeling-Llama-3.2-1B-bnb-4bit"
inferencer = LlamaInference(model_path)
response = inferencer.generate_response("Your text here", stream=False)
Output Format
The model generates hierarchical topics in the format:
Domain > High-Level Topic > Specific Topic
Example:
Input: """Machine learning (ML) is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.[1] Advances in the field of deep learning have allowed neural networks to surpass many previous approaches in performance.[2] ML finds application in many fields, including natural language processing, computer vision, speech recognition, email filtering, agriculture, and medicine.[3][4] The application of ML to business problems is known as predictive analytics. Statistics and mathematical optimization (mathematical programming) methods comprise the foundations of machine learning. Data mining is a related field of study, focusing on exploratory data analysis (EDA) via unsupervised learning.[6][7] From a theoretical viewpoint, probably approximately correct (PAC) learning provides a framework for describing machine learning."""
Output: "Machine Learning > Definition > Overview"
License
This project is released under LLaMA-3.2 License.
Citation
If you use this model or dataset in your research, please cite:
@misc{dynamic-topic-llama,
author = {Aman Priyanshu},
title = {Dynamic Topic Modeling Llama 3.2 1B},
year = {2024},
publisher = {HuggingFace}
}