• Fine-tuned the LLM base of two multi-modal large language models (7B and 1B parameters) with 60k data samples on 32 V100 GPUs on phone screen navigation tasks, achieving 75% accuracy in multi-step predictions of music-app-related actions
• Innovated an automatic data generation pipeline that created 2k UI screen data samples for fine-tuning tasks, leveraging opencv to synthesize android screens and in-app designs, surpassing traditional web scraping and manual labeling methods 20 times in speed
• Iterated the fine-tuning procedures from merely generating action predictions to also producing Chain-of-Thought data including observations, action logic, and predicted results, validating ideas with 5+ similar pipelines like ScreenAgent and MobileAgent
• Took full ownership of the automatic generation of Chain-of-Thought data as training output for a 20B MLLM model, running GPT4V to generate 10k instances in 20+ parallel subprocesses, improving model accuracy on multiple benchmarks by 15%
• Creatively implemented another data pipeline that randomly combined 15k authentic screenshots into 3k simulated screen operation data, using ui-animator to navigate Android apps and GPT4V to generate corresponding metadata

• Led a team of three to compile the first benchmark to test diffusion models’ ability to generate images grounded in common sense
• Designed a dataset across 6 common sense categories containing 200 pairs of image prompts, following up with an automatic evaluation process to use CLIP to compute cosine similarity between generated images and ground truth
• Constructed an automatic data pipeline from few-shot prompting GPT-4 to generate 200+ prompts per batch, to serializing and transferring prompts to remote GPU pools as model inputs, to generating and evaluating 4 variations of each prompt in parallel
• Innovated 2 automatic scoring metrics on 4 SOTA models that only diverges from human evaluations by 3%, revealing major shortcomings in new Dall-E and Stable Diffusion models including the lack of basic physical and spatial common sense

• Deployed a philosophy chatbot platform with Next.js and Vercel, enabling users to create and chat with AI philosophers by storing the models in a SQL-based Prisma database and routing users’ queries to GPT-4’s API and a vector database
• Created an ample external knowledge database by recursively scraped 500+ web pages from Stanford Encyclopedia of Philosophy with requests and BeautifulSoup, splitting the texts into 46000+ chunks of data and upserting them to a Pinecone vector database
• Improved the accuracy of chatbots’ responses using similarity search to return the top 3 text chunks most relevant to the query within 0-1 second and embedding them in GPT prompts, surpassing native GPT’s user satisfaction rate as reported by 10+ users
• Enabled chatbots to have long-term memories about user conversations by storing all chat messages into Redis and retrieving the 30+ most recent messages in realtime, preventing chatbots’ from memory loss due to recurrent API calls

• Created a decoder-only transformer with 16 million tokens capable of generating text in any given author’s writing style
• Preprocessed the complete corpus of Shakespeare with 1.2 million text tokens, chunking the text into sequences of 32-character tokens and transforming them into embeddings of size 256 to serve as input to the transformer model
• Manually implemented 6 decoder blocks of multi-head attention layers and feed-forward layers to enable the model to understand semantic relations between tokens, adding residual connections and pre-layer normalization to prevent vanishing gradient
• Successfully achieved a cross-entropy loss of 0.9 on the training set and 1.5 on the validation set after 4k iterations of mini-batch gradient descent using the Adam optimizer, approaching the authentic writing styles of the target author

REDPhilosophy content creator

• Innovated popular philosophical content for the general public, acquiring 25,000 followers within 5 months and surpassing 93% of channels in growth speed
• Produced animations, articles, and cartoons every two days, acquiring over 1,000 interactions across social media per post