Thesis glossing experiments

This repository contains the experiment code for the MSc thesis:

Prompting and Fine-tuning Large Language Models for Interlinear Glossing

The code is organized by model family and covers the training, inference, retrieval, and evaluation pipelines used in the thesis.

Repository structure

• glosslm/ — GlossLM fine-tuning, inference, prompt-variant evaluation, and WAV2Gloss cross-dataset evaluation.
• polygloss/ — PolyGloss inference and evaluation on the PolyGloss dataset.
• llama/ — LLaMA prompting, chrF++ retrieval, LoRA fine-tuning, cached few-shot fine-tuning, and PolyGloss cross-dataset evaluation.

What is not included

The repository intentionally excludes:

• model checkpoints and LoRA adapters
• cached datasets and retrieved-example caches
• prediction files and evaluation outputs
• logs and SLURM job files
• local virtual environments
• Hugging Face caches and tokens

These files are either too large, environment-specific, or should not be committed.

Data and models

The experiments rely on external datasets and models, including:

• GlossLM corpus / GlossLM split data
• PolyGloss corpus
• WAV2Gloss text data
• meta-llama/Llama-3.1-8B-Instruct
• GlossLM and PolyGloss checkpoints

Local paths should be supplied by the user through config files or command-line arguments.

Reproducibility note

The code is intended to document and reproduce the thesis experiment pipelines. Exact reproduction requires access to the same model checkpoints, dataset versions, and local compute environment.