You can deploy this model and start earning money today!

anshulmago1/
ap-rowa-thesis-grader-qwen3-0.6b

Name: ap-rowa-thesis-grader-qwen3-0.6b
Availability: InStock

Views

No views yet

Row A Thesis-Point Grader — Qwen3-0.6B LoRA

A LoRA adapter for Qwen/Qwen3-0.6B that grades the AP World History LEQ/DBQ Row A (thesis/claim) point. Input: an LEQ/DBQ prompt + a candidate thesis. Output: {"point": 0|1, "reason": "..."} matching the official College Board Row A decision.

Behavior spec

Awards a minimal-but-defensible thesis (a plain claim with any one line of reasoning), denies restatements / off-topic / non-defensible claims, and does not import higher-row criteria ("evaluate the extent," analytical complexity).

Results (71 held-out real College Board student theses)

model	agreement	Cohen's κ	false-deny	false-award
base Qwen3-0.6B (untuned)	79%	0.10	0	14
this adapter (tuned)	80%	0.54	13	1
gpt-4o (hardened prompt)	75%	0.39	14	4
gpt-4o (decomposed)	82%	0.49	8	5

Fine-tuning lifts the base model from κ 0.10 → 0.54, beating every prompted gpt-4o configuration at ~1/1000th the size.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
m = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B")
m = PeftModel.from_pretrained(m, "<this-repo>")
# Use the compact system prompt from src/rowa/rubric.py (grader_system("compact")).

Trained with QLoRA/LoRA (TRL + PEFT). See the dataset card and Brainlift.md for method.

anshulmago1/ap-rowa-thesis-grader-qwen3-0.6b

Row A Thesis-Point Grader — Qwen3-0.6B LoRA

Behavior spec

Results (71 held-out real College Board student theses)

Usage

anshulmago1/
ap-rowa-thesis-grader-qwen3-0.6b