Skip to content

Latest commit

 

History

History
66 lines (40 loc) · 2.49 KB

File metadata and controls

66 lines (40 loc) · 2.49 KB

LLM Classification Finetuning

Ensembled Pipeline Inference for Human Preference Classification


Notebook Author


Open in Kaggle


Notebook in this folder

llm_classification_inference.ipynb


Competitions  ·  Achievements  ·  Courses  ·  Kaggle Profile


The problem

Given a prompt and two model responses, predict which one a human preferred. The difficulty is positional bias: a model that reads response A first will tend to favour it, and that preference has nothing to do with the responses themselves.

What is here

The countermeasure is to run the same data twice, once in each order, and average the results. Each pass uses a different backbone, so the ensemble also averages over architecture.

Stage What it does
Data structuring Builds two views of the test set, one standard and one with Response A and Response B interchanged
Gemma-2 pass Runs Gemma-2-9B with manual layer allocation across two GPUs and a variable-length collator
Llama-3 pass Repeats the partitioning for Llama architecture dimensions, run against the swapped view
Ensemble Averages the two passes, cancelling the order effect that either pass would carry alone

Important

Inference runs with the internet disabled. The model weights and the human_pref helper module are attached as Kaggle datasets rather than installed, and the notebook expects them to be present in the environment.

Stack  ·  torch transformers xformers pandas numpy


Amey Thakur  ·  Kaggle  ·  GitHub  ·  ORCID


↑ Back to top  ·  ← Repository home