Fine-tuning small language models for named entity recognition on Tweebank-NER. Based on the psychological concept of task demands, we propose a "high-demand" and "low-demand" evaluation method / data format and compare their performance.
| File | Description |
|---|---|
llama.ipynb |
Pipeline for Llama 3.1 8B |
mistral.ipynb |
Pipeline for Mistral v0.3 7B |
qwen.ipynb |
Pipeline for Qwen 3 4B |
f1.py |
Scoring script: computes overall and per-entity-class precision/recall/F1 from model predictions |
Pipelines are adapted from Unsloth notebooks. Each contains code for data formatting, QLoRA fine-tuning, generation, and evaluation.