formrecap-gemma-2b-lora

LoRA adapter fine-tuned on synthetic form abandonment event traces for 6-class classification.

Base Model

google/gemma-2b-it

Task

Classifies form interaction event traces into one of six abandonment reasons:

Code	Class	Description
1	validation_error	User hit a field error they couldn't resolve
2	distraction	User task-switched away
3	comparison_shopping	Browsing, not committing
4	accidental_exit	Closed tab / back button by mistake
5	bot	Automated non-human interaction
6	committed_leave	Intentionally chose not to complete

Training

• Method: QLoRA (NF4 4-bit) + LoRA (r=16, alpha=32, DoRA=True)
• Data: 884 synthetic examples generated with Claude Sonnet, stratified 80/10/10 split
• Hardware: Modal L4 GPU, ~30 minutes, ~$0.40
• Framework: HuggingFace PEFT + TRL

Evaluation (52 hand-labeled real test examples)

Metric	Value
Macro-F1	0.916
ECE (logprob)	0.103

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("google/gemma-2b-it")
model = PeftModel.from_pretrained(base, "jasonm4130/formrecap-gemma-2b-lora")
tokenizer = AutoTokenizer.from_pretrained("jasonm4130/formrecap-gemma-2b-lora")

Source

jasonm4130/formrecap-lora-classifier