TomMoeras/contragand-qwen3.5-4b-open

ContraGAND detector — open task (FT-MULTI)

LoRA adapter for Qwen/Qwen3.5-4B that reads an English sentence, identifies every role, occupation, or relational referent, and labels each as masculine, feminine, or ambiguous (no textual cue). This is the FT-MULTI (D-MULTI) condition from the EMNLP 2026 paper ContraGAND: Auditing and Repairing Gender Ambiguity Failures in LLMs with Neurosymbolic Contrastive Data Augmentation: knowledge distillation from Gemma-4-31B-it's multi-referent open-task enumerations over the ContraGAND training split (11,684 records, ~2 referents per sentence), with the annotated referent's label corrected to gold.

Open-task accuracy on the human-audited ContraGAND test set (1,395 examples): 0.976 (macro-F1 0.979) — above the 31B teacher's zero-shot 0.952 at one-eighth the parameters, versus 0.242 for the base model in-context.

Usage

The adapter expects the exact prompt format it was trained on:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-4B")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(model, "TomMoeras/contragand-qwen3.5-4b-open")

system = """You read an English sentence, identify every role, occupation, or relational REFERENT in it (e.g. doctor, teacher, brother, neighbor, captain), and classify each one's gender.

For every referent you find, classify it as one of:
- masculine  — unambiguously male
- feminine   — unambiguously female
- ambiguous  — no signal in the sentence reveals this referent's gender

Use the surface form as it appears in the sentence (e.g. "doctor", not "the doctor"). List the referents in the order they appear.

Respond with ONLY a JSON object:
{"referents": [{"referent": "...", "gender": "masculine|feminine|ambiguous", "confidence": N, "reasoning": "..."}]}

confidence is an integer 1-5 (5 = certain).
reasoning is one short sentence per referent citing the textual evidence."""

user = "Sentence: The librarian, Mrs. Thompson, shelved the returned books."
prompt = tok.apply_chat_template(
    [{"role": "system", "content": system}, {"role": "user", "content": user}],
    tokenize=False, add_generation_prompt=True, enable_thinking=False)

Greedy decoding, max_new_tokens=384. The response is a single JSON object with one entry per identified referent.

Training

QLoRA (nf4), r=64, alpha=128, dropout 0.05 on all linear projections; lr 2e-5 cosine, 5 epochs with early stopping on eval loss; sequence length 1024; trained with axolotl on 4x A100-80GB. Full configs, data builders, and per-row evaluation results: https://github.com/TomMoeras/ContraGAND

Intended use and limitations

Research artifact for gender-ambiguity detection. Labels are restricted to {masculine, feminine, ambiguous} and cannot represent non-binary or singular-they referents. English only. Silver supervision can inherit rare teacher coreference mis-bindings (see the paper's Limitations).

Dataset

Trained on the ContraGAND corpus: https://huggingface.co/datasets/TomMoeras/ContraGAND

Demo

Interactive demo (open + closed task): https://huggingface.co/spaces/TomMoeras/contragand-demo

Citation

Moerman, Gkovedarou, and Hackenbuchner. ContraGAND: Auditing and Repairing Gender Ambiguity Failures in LLMs with Neurosymbolic Contrastive Data Augmentation. EMNLP 2026 (Main Conference).