DAT VO DINH

LLM Finetuning

4/2024 - Now

• Tools: Transformers, Unsloth, PEFT, Trl
• Finetune open sources LLM model such as Llama-3, Mistral to perform multiple task such as:
• Function calling: enable LLM to use external functions and tools.
• Instruction tuned: generate detail answers based on the given instruction.
• Structured output: LLM return structured output such as JSON, XML...

RAG Chatbot

3/2024 - Now

• Tools: Ollama, LlamaIndex, ChromaDB, Gradio, Docker, FastAPI, Ngrok
• Build a fully local Chatbot with advanced RAG pipeline that can retrieve information from multiple documents based on user queries and providing precise answers.
• Pull and use any open sources model from Huggingface and Ollama with different quantization.
• Design a simple web interface so users can customize LLM models and input documents.
• Implement Docker for streamlined deployment and management of the Chatbot system

Stable Diffusion from scratch

2/2024

• Tools: Pytorch, Lightning, WandB, Gradio
• Implemented Latent Diffusion Model (Stable Diffusion) from scratch for education purpose, include full training and tracking experiment pipeline using Pytorch Lightning and Wandb.
• Build a simple website to visualize the inference of the diffusion model.

Freelance Quantitative Researcher

10/2023 - 01/2024

Taiwan

• Technologies: Pytorch, Lightning, WandB, Transformers, Polars, Amazon S3
• Design and improve the reward function for a reinforcement learning-based trading environment to mimic real-world
trading scenarios.
• Developing an RL agent combine with current SOTA algorithms such as Transformers, surpass the performance of
traditional ensemble models such as XGBoost, Gradient Boosting,... through backtest evaluation.

Part time AI Researcher

5/2022 - 7/2023

Ha Noi, Vietnam

• Technologies: Pytorch, Numpy, Numba, Gynasium, Selenium.
• Researched and implemented 2 deep reinforcement learning algorithms, successfully surpass 95% of the environments
in the system by competing with other decision-making agents.
• Developed three gym-based multiplayer environments for researching algorithms, with optimization achieved using
Numpy and Numba JIT, reaching the speed upto 1000 games per second.
• Research about Japanese stock market, crawl data from 3 different sources: Minkabu, MarketWatch and YahooJP
and cross-check to ensure the credibility of sources.