AgriFair-GRAFT-adapters

AgriFair GRAFT adapters

LoRA adapters from GRAFT (Gradient-Ranked Adapter Fairness Targeting), a study of difference-aware fairness in agricultural language models, grounded in the Indian Agriculture Census 2015-16 through the AgriFair benchmark.

The method locates where a model encodes gap erasure, answering "Roughly equal" where the census records a real difference, using Integrated Gradients. It places LoRA adapters on those layers alone with rank proportional to attribution, and trains them with a condition-adaptive, rationale-aware objective.

This repository holds 132 adapters so that every number in the paper can be recomputed rather than taken on trust.

Layout

<tier>/<method>/seed_<n>/     adapter_config.json, adapter_model.safetensors
manifest.json                 base model, revision, rank range and trainable share per adapter

Methods present: ablation_no_condition_adaptive, ablation_no_rationale_loss, ablation_placement_random, ablation_placement_random_draw2, ablation_placement_random_draw3, ablation_placement_uniform, ablation_rank_sweep_16, ablation_rank_sweep_32, ablation_rank_sweep_8, baseline_dart, baseline_fairnet, baseline_igu_lora, baseline_lftf, baseline_pedal, baseline_regift, graft_proposed, graft_proposed_loao_gender, graft_proposed_loao_landholding, graft_proposed_loao_social_group, graft_proposed_qlora_nf4, reference_vanilla_lora, reference_vanilla_lora_loao_gender, reference_vanilla_lora_loao_landholding, reference_vanilla_lora_loao_social_group, reference_vanilla_qlora

Using one

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen3-4B-Instruct-2507"
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "Debk/AgriFair-GRAFT-adapters", subfolder="broad-instruct/ablation_no_condition_adaptive/seed_42")
tokenizer = AutoTokenizer.from_pretrained(base)

manifest.json gives the exact base model and revision each adapter was trained against. Loading an adapter onto a different base, or a different revision, is not a valid comparison.

Scope and limitations

These adapters are trained to change behaviour on questions about caste, gender and landholding inequality in Indian agriculture. They are a research artefact:

• Trained and evaluated in English only, on a templated benchmark.
• AgriFair concedes 0.872 to a predictor that never reads the state name, so any accuracy must be read against that ceiling. The study reports a structure-novel test slice for exactly this reason.
• The gender axis has 180 items, since the census reports gender only at national level, so gender results carry wide intervals.
• Not validated for deployment in any advisory system, and not a substitute for agronomic or policy advice.

Citation

Cite the All India Report on Agriculture Census 2015-16 (Department of Agriculture and Farmers Welfare, Government of India) as the source of every AgriFacts answer, and KisanVaani/agriculture-qa-english-only for the AgriAdvice base queries.