Downloads · 30 days
14
33% of all-time downloads
Stereotypes-in-LLMs/qwen3.5-9b-hiring-debias-sft-en
qwen3.5-9b-hiring-debias-sft-en is a machine learning model from Stereotypes-in-LLMs. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
LoRA adapter that makes Qwen/Qwen3.5-9B decide hiring cases without letting a protected attribute (military status, gender, religion) change the verdict. Trained on counterfactually invariant targets; audited on a hel…
Downloads · 30 days
14
33% of all-time downloads
All-time downloads
43
Public
Repo size
253 MB
Likes
0
Public
Click a slice to open those files.
.safetensors233 MB · 92%
From the Hugging Face model README
LoRA adapter that makes Qwen/Qwen3.5-9B decide hiring cases without letting a protected attribute (military status, gender, religion) change the verdict. Trained on counterfactually invariant targets; audited on a held-out counterfactual benchmark of 31,050 decisions.
Qwen3.5-9B--en--sft--adapter--en_only (every decision, reproducible without a GPU)A counterfactual set is one candidate–job pair evaluated with every attribute variant; it is unstable when the decision is not the same across them — the attribute alone tipped it. Paired with the same sets at baseline, on matched variants, 900 sets:
| Base model | This adapter | |
|---|---|---|
| Unstable sets | 17.3% | 7.2% |
| Change | -10.1 pp [-14.4, -5.8] | |
| Sets fixed : broken | 136 : 45 | |
| Utility (agreement with the attribute-free reference) | 78.9% | 82.5% |
| Rationales naming the attribute | 1.1% | 0.2% |
Per protected group, all significant after Benjamini–Hochberg correction:
| Group | Base unstable % | Adapter unstable % | Δ pp | p (FDR) |
|---|---|---|---|---|
| gender | 7.7 | 4.1 | -3.6 | 1e-03 |
| military status | 10.9 | 3.8 | -7.1 | 1e-08 |
| religion | 8.1 | 2.8 | -5.3 | 1e-06 |
Fine-tuning moves the hire rate (14.5% → 27.3%), and instability depends on where the decision threshold sits. Sweeping that threshold on each model's own hire-vs-reject margin over all 31,050 audited prompts:
| At a hire rate of 27.3% | Unstable sets % |
|---|---|
| base model, moved to that rate | 13.0 |
| this adapter | 3.5 |
At equal hire rate the adapter is 3.7× more consistent; -106% of the raw gain is attributable to the shift. The adapter's curve lies below the base model's at every operating point.
Do not serve this adapter through vLLM's LoRA path. For these architectures that path does not reproduce the trained model: on the same prompts, HuggingFace + PEFT and vLLM agree on 95–99% of the base model's decisions but only 38–83% of the adapter's. Fold the adapter into the weights and serve the result as an ordinary checkpoint:
from peft import PeftModel
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", dtype="bfloat16")
model = PeftModel.from_pretrained(model, "Stereotypes-in-LLMs/qwen3.5-9b-hiring-debias-sft-en").merge_and_unload()
model.save_pretrained("merged") # then: vllm serve merged
The study's audits all ran on merged weights (scripts/merge_adapter.py).
The paper is in preparation; until then cite the repository above and the Djinni Recruitment Dataset (Drushchak & Romanyshyn, 2024).