the entire day was spent on the abstract. i'm working on two projects. outcome prediction and embedding of brain tumor notes. i added other embedding models to have a more abstract-worthy benchmark. i had to make sure my finetuning works better. if it doesn't, the entire abstract is pointless. it's my first ever abstract, i felt immense pressure to succeed. all my metrics had to be bolded. i was tweaking everything, from running finetuning again with more epochs and a different learning rate, to cleaning the text, to changing the T-SNE parameters, to adding stratified sampling for knn classification, and adding the task specific LoRA adaptors, changing the task type, and finally settling on classification. it's interesting that the classification task going against my hypothesis. i have to understand deeper what this is doing. i've just been turning the knobs, crossing my fingers, and hoping for the numbers to go up.

the outcome prediction was even messier, OSS models like llama and mistral are very finnicky, a slight change in the prompts, the presence of one sentence, can steer the model into the ground, or towards the jackpot. my expectations were set too high for the results, i have to be prepared to be surprised, to not come with any presuppositions in research. test everything for yourself. your intuition can be wrong.

this is the prompt that i used.

OUTCOME_PROMPT_TEMPLATE = """Your task is to predict the 5-year survival outlook for the following cancer patient based on an aggregation of clinical notes up to {date_span_days} days post-diagnosis.

Here are the clinical notes:
{clinical_note}

Please analyze these clinical notes and extract the most pertinent information in the form of 5 key factors. Based on this and the rest of the notes, classify the patient's 5-year survival outlook. Respond with either 'POSITIVE' (if the patient is likely to survive beyond 5 years) or 'NEGATIVE' (if the patient is unlikely to survive beyond 5 years).
"""

SYSTEM_MESSAGE = """You are an oncologist specializing in head and neck cancers, tasked with predicting survival outcomes for patients."""