renhehuang/fair-resume-job-matcher-lora

fairness_lora

Fairness-aware LoRA adapter for resume–job matching built on top of BAAI/bge-large-en-v1.5.

This adapter was trained with adversarial debiasing and multi-task objectives to reduce group disparities while maintaining utility.

Usage

from transformers import AutoTokenizer, AutoModel
from peft import PeftModel

BASE = "BAAI/bge-large-en-v1.5"
ADAPTER = "fairness_lora"

tokenizer = AutoTokenizer.from_pretrained(BASE)
base_model = AutoModel.from_pretrained(BASE)
model = PeftModel.from_pretrained(base_model, ADAPTER)
model.eval()

# Encode two texts and compute cosine
import torch
import torch.nn.functional as F

def encode(text):
    enc = tokenizer(text, return_tensors='pt', truncation=True, padding=True, max_length=256)
    with torch.no_grad():
        out = model(**enc)
        hidden = out.last_hidden_state
        emb = F.normalize(hidden.mean(dim=1), p=2, dim=1)
    return emb

r = "Software engineer with Python experience."
j = "Hiring backend Python developer."
cos = (encode(r) * encode(j)).sum(dim=1).item()
prob = torch.sigmoid(torch.tensor(cos)).item()
print({"cosine": cos, "prob": prob})

Training notes

• Base model: BAAI/bge-large-en-v1.5
• Technique: LoRA + adversarial debiasing + multi-task attribute prediction
• Objective: Lower demographic parity / equalized odds gaps while preserving accuracy/AUC

License

• Please ensure the license of the base model BAAI/bge-large-en-v1.5 allows derivative adapters.
• Provide your dataset and usage terms accordingly.

Framework versions

• PEFT 0.17.1