Circuit-Full cross-encoder fine-tuned on AgXQA agricultural dataset. Attention + MLP in IE-ranked layers 8-11 (4.73M params). Initialized from SciFact-fine-tuned JAYADIR/mdts-circuit-full-bm25. Best efficiency score among all strategies.

Training

Stage	Dataset	Description
Stage 1	SciFact (BEIR)	Circuit fine-tuning with BM25 hard negatives
Stage 2	AgXQA	Domain adaptation to agricultural extension QA

Results on AgXQA Test Set (NDCG@10)

Strategy	Params	Zero-Shot	AgXQA FT	Delta	MRR@10
BM25 baseline	—	0.8533	—	—	0.8208
Vanilla (no FT)	—	0.9124	—	—	0.8907
A: Circuit MLP-only	2.36M	0.9123	0.9151	+0.0028	0.8945
B: Last-4 Layers	7.10M	0.9070	0.9135	+0.0065	0.8937
C: Full Fine-Tuning	33.36M	0.9020	0.9231	+0.0211	0.9051
D: Circuit-Full (BM25)	4.73M	0.9094	0.9174	+0.0080	0.8976

This model: D: Circuit-Full (BM25) | Params: 4.73M | NDCG@10: 0.9174 | MRR@10: 0.8976

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("JAYADIR/mdts-agxqa-circuit-full-bm25")
model = AutoModelForSequenceClassification.from_pretrained("JAYADIR/mdts-agxqa-circuit-full-bm25")

query   = "What is the best irrigation method for wheat in semi-arid regions?"
passage = "Drip irrigation is highly efficient for wheat cultivation in semi-arid areas."

inputs = tokenizer(query, passage, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
    score = model(**inputs).logits.squeeze().item()
print(f"Relevance score: {score:.4f}")