Stereotypes-in-LLMs/hiring-bias-mitigation-responses

Hiring-bias mitigation — model responses

Every response produced in the mitigation study of LLM hiring decisions: 64 runs, 2,782,350 responses, from 5 open-weight models in English and Ukrainian, at baseline and under each mitigation family (baseline, embedding, prompt, scrub, sft). Each run is one subset.

• All released artifacts: the Hiring Bias Mitigation collection.
• Training data of the fine-tuned runs: hiring-bias-mitigation-synthetic-data.
• Code, configs, full results and findings: Stereotypes-in-LLMs/hiring_bias_mitigation — see reports/RESULTS.md and reports/MITIGATION_FINDINGS.md.
• The audit study this extends (benchmark, attributes, injection templates, and its own responses): the Stereotypes-in-LLMs/hiring-analyses-* datasets, e.g. hiring-analyses-baseline-en.
• Underlying CVs and job descriptions: the Djinni Recruitment Dataset (candidate profiles, job descriptions).

Fine-tuned runs are included. The six SFT cells (Qwen3.5-4B/9B and LAPA-12B, English and Ukrainian) plus one decision-weighted probe, all audited from merged weights: vLLM's LoRA path does not reproduce these adapters (agreement with HuggingFace + PEFT is 38–83% depending on the architecture, against 95–99% for the base model), so every adapter is folded into the weights before serving. Preference-optimisation (DPO) runs are internal probes and are not published.

Design

Each benchmark job–CV pair is shown to a model many times, identical except for one injected protected attribute (military status, gender or religion), stated explicitly or implicitly (a first-person sentence), plus an attribute-free control. Rows sharing a group_id form one counterfactual set: a fair screener gives them all the same decision. Decoding is greedy (temperature 0, seed 42), with thinking mode off.

Family	What changes
baseline	nothing — the unmitigated model
prompt	the instruction (e.g. structured_rubric, ignore_personal_info)
scrub	the attribute is removed from the input first (lexical rules or an LLM)
embedding	LEACE concept erasure of the attribute at one hidden layer

Loading

from datasets import load_dataset
ds = load_dataset("Stereotypes-in-LLMs/hiring-bias-mitigation-responses", "Qwen3.5-9B--en--prompt--structured_rubric", split="test")

To re-score a run without a GPU, download runs/<run>.parquet into the code repository's outputs/raw/ and run python scripts/run_audit.py --config <its config> --score-only.

Columns

Column	Meaning
pair_id, candidate_id, job_id	benchmark identifiers
cv, job_description, job_position, lang	the inputs, verbatim
protected_group, protected_group_label, protected_attr	the injected attribute
condition	explicit, implicit or attr_free
implicit_injection	the first-person sentence used in the implicit condition
profile_rendered	the candidate profile exactly as the model saw it
group_id	counterfactual set key
raw_output	the model's response, unmodified
decision, feedback, outcome, raw_decision	the parsed decision and rationale, and whether parsing succeeded
reference_decision, reference_feedback	the attribute-free GPT-4o reference from the audit study
feedback_similarity	cosine similarity of feedback to the reference rationale

runs/<run>.meta.json holds the model, decoding settings, seed, mitigation block and runtime; results/<run>.json the scored summary (disparity tests, refusal and parse rates, utility). tables/ holds the study-level set-stability comparisons on matched variants.

Runs

Utility is agreement with the attribute-free reference decision. Runs marked not usable are published for completeness and excluded from every analysis.

Run	Model	Lang	Family	Strategy	Groups	Conditions	Rows	Parse fail %	Utility %	Usable
Qwen3.5-4B--en--baseline	Qwen3.5-4B	en	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.0	80.6	yes
Qwen3.5-4B--en--prompt--counterfactual_invariance	Qwen3.5-4B	en	prompt	counterfactual_invariance	military_status, religion	explicit, implicit, attr_free	13,050	0.0	75.3	yes
Qwen3.5-4B--en--prompt--fairness_constitution	Qwen3.5-4B	en	prompt	fairness_constitution	military_status, religion	explicit, implicit, attr_free	13,050	0.0	79.9	yes
Qwen3.5-4B--en--prompt--ignore_personal_info	Qwen3.5-4B	en	prompt	ignore_personal_info	military_status, religion	explicit, implicit, attr_free	13,050	0.0	80.9	yes
Qwen3.5-4B--en--prompt--reasoning	Qwen3.5-4B	en	prompt	reasoning	military_status, religion	explicit, implicit, attr_free	13,050	0.0	79.3	yes
Qwen3.5-4B--en--prompt--recruiter_guidelines	Qwen3.5-4B	en	prompt	recruiter_guidelines	military_status, religion	explicit, implicit, attr_free	13,050	0.0	78.1	yes
Qwen3.5-4B--en--prompt--second_pass_verification	Qwen3.5-4B	en	prompt	second_pass_verification	military_status, religion	explicit, implicit, attr_free	13,050	0.0	68.4	yes
Qwen3.5-4B--en--prompt--structured_rubric	Qwen3.5-4B	en	prompt	structured_rubric	military_status, religion	explicit, implicit, attr_free	13,050	0.1	74.2	yes
Qwen3.5-4B--en--prompt--zero_shot_cot	Qwen3.5-4B	en	prompt	zero_shot_cot	military_status, religion	explicit, implicit, attr_free	13,050	0.0	80.0	yes
Qwen3.5-4B--en--scrub--lexical	Qwen3.5-4B	en	scrub	lexical	military_status, religion	explicit, implicit, attr_free	13,050	0.0	83.4	yes
Qwen3.5-4B--en--scrub--llm	Qwen3.5-4B	en	scrub	llm	military_status, religion	explicit, implicit, attr_free	13,050	0.0	81.5	yes
Qwen3.5-4B--en--embedding--leace	Qwen3.5-4B	en	embedding	leace (layer 16)	military_status, religion	explicit, implicit, attr_free	13,050	19.0	80.9	yes
Qwen3.5-4B--en--sft--adapter--en_only	Qwen3.5-4B	en	sft	qwen3.5-4b_en_only	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	82.0	yes
Qwen3.5-4B--uk--baseline	Qwen3.5-4B	uk	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.0	74.8	yes
Qwen3.5-4B--uk--prompt--counterfactual_invariance	Qwen3.5-4B	uk	prompt	counterfactual_invariance	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	73.6	yes
Qwen3.5-4B--uk--prompt--fairness_constitution	Qwen3.5-4B	uk	prompt	fairness_constitution	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	74.7	yes
Qwen3.5-4B--uk--prompt--ignore_personal_info	Qwen3.5-4B	uk	prompt	ignore_personal_info	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	75.2	yes
Qwen3.5-4B--uk--prompt--reasoning	Qwen3.5-4B	uk	prompt	reasoning	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	74.0	yes
Qwen3.5-4B--uk--prompt--recruiter_guidelines	Qwen3.5-4B	uk	prompt	recruiter_guidelines	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	74.1	yes
Qwen3.5-4B--uk--prompt--second_pass_verification	Qwen3.5-4B	uk	prompt	second_pass_verification	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	74.0	yes
Qwen3.5-4B--uk--prompt--structured_rubric	Qwen3.5-4B	uk	prompt	structured_rubric	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.3	69.6	yes
Qwen3.5-4B--uk--prompt--zero_shot_cot	Qwen3.5-4B	uk	prompt	zero_shot_cot	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	73.0	yes
Qwen3.5-4B--uk--scrub--lexical	Qwen3.5-4B	uk	scrub	lexical	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	75.4	yes
Qwen3.5-4B--uk--scrub--llm	Qwen3.5-4B	uk	scrub	llm	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	74.5	yes
Qwen3.5-4B--uk--embedding--leace	Qwen3.5-4B	uk	embedding	leace (layer 16)	gender, military_status, religion	explicit, implicit, attr_free	31,050	100.0	--	no — 100.0% of responses could not be parsed (only 4 of 31,050 decided)
Qwen3.5-4B--uk--sft--adapter--uk_only	Qwen3.5-4B	uk	sft	qwen3.5-4b_uk_only	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	77.8	yes
Qwen3.5-9B--en--baseline	Qwen3.5-9B	en	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.0	78.0	yes
Qwen3.5-9B--en--prompt--counterfactual_invariance	Qwen3.5-9B	en	prompt	counterfactual_invariance	military_status	implicit, attr_free	2,700	0.0	74.9	yes
Qwen3.5-9B--en--prompt--fairness_constitution	Qwen3.5-9B	en	prompt	fairness_constitution	military_status	implicit, attr_free	2,700	0.0	83.9	yes
Qwen3.5-9B--en--prompt--ignore_personal_info	Qwen3.5-9B	en	prompt	ignore_personal_info	military_status	implicit, attr_free	2,700	0.0	78.1	yes
Qwen3.5-9B--en--prompt--ignore_personal_info--fullscope	Qwen3.5-9B	en	prompt	ignore_personal_info	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	80.0	yes
Qwen3.5-9B--en--prompt--reasoning	Qwen3.5-9B	en	prompt	reasoning	military_status	implicit, attr_free	2,700	0.0	74.4	yes
Qwen3.5-9B--en--prompt--recruiter_guidelines	Qwen3.5-9B	en	prompt	recruiter_guidelines	military_status	implicit, attr_free	2,700	0.0	74.5	yes
Qwen3.5-9B--en--prompt--second_pass_verification	Qwen3.5-9B	en	prompt	second_pass_verification	military_status	implicit, attr_free	2,700	0.0	77.5	yes
Qwen3.5-9B--en--prompt--second_pass_verification--fullscope	Qwen3.5-9B	en	prompt	second_pass_verification	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	79.8	yes
Qwen3.5-9B--en--prompt--structured_rubric	Qwen3.5-9B	en	prompt	structured_rubric	military_status	implicit, attr_free	2,700	0.0	70.0	yes
Qwen3.5-9B--en--prompt--structured_rubric--fullscope	Qwen3.5-9B	en	prompt	structured_rubric	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	72.6	yes
Qwen3.5-9B--en--prompt--zero_shot_cot	Qwen3.5-9B	en	prompt	zero_shot_cot	military_status	implicit, attr_free	2,700	0.0	76.4	yes
Qwen3.5-9B--en--scrub--lexical	Qwen3.5-9B	en	scrub	lexical	military_status	implicit, attr_free	2,700	0.0	80.3	yes
Qwen3.5-9B--en--scrub--llm	Qwen3.5-9B	en	scrub	llm	military_status	implicit, attr_free	2,700	0.0	80.8	yes
Qwen3.5-9B--en--embedding--leace	Qwen3.5-9B	en	embedding	leace (layer 16)	military_status	implicit, attr_free	2,700	0.0	73.1	yes
Qwen3.5-9B--en--sft--adapter--en_only	Qwen3.5-9B	en	sft	qwen3.5-9b_en_only	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	82.5	yes
Qwen3.5-9B--en--sft--adapter--en_only_v2_ckpt250	Qwen3.5-9B	en	sft	checkpoint-250	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	81.4	yes
Qwen3.5-9B--uk--baseline	Qwen3.5-9B	uk	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.0	79.5	yes
Qwen3.5-9B--uk--prompt--counterfactual_invariance	Qwen3.5-9B	uk	prompt	counterfactual_invariance	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	79.9	yes
Qwen3.5-9B--uk--prompt--fairness_constitution	Qwen3.5-9B	uk	prompt	fairness_constitution	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	79.9	yes
Qwen3.5-9B--uk--prompt--ignore_personal_info	Qwen3.5-9B	uk	prompt	ignore_personal_info	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	80.9	yes
Qwen3.5-9B--uk--prompt--reasoning	Qwen3.5-9B	uk	prompt	reasoning	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	80.2	yes
Qwen3.5-9B--uk--prompt--recruiter_guidelines	Qwen3.5-9B	uk	prompt	recruiter_guidelines	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	80.2	yes
Qwen3.5-9B--uk--prompt--second_pass_verification	Qwen3.5-9B	uk	prompt	second_pass_verification	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	76.7	yes
Qwen3.5-9B--uk--prompt--structured_rubric	Qwen3.5-9B	uk	prompt	structured_rubric	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	77.1	yes
Qwen3.5-9B--uk--prompt--zero_shot_cot	Qwen3.5-9B	uk	prompt	zero_shot_cot	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	81.0	yes
Qwen3.5-9B--uk--scrub--lexical	Qwen3.5-9B	uk	scrub	lexical	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	81.7	yes
Qwen3.5-9B--uk--scrub--llm	Qwen3.5-9B	uk	scrub	llm	gender, military_status, religion	explicit, implicit, attr_free	31,050	0.0	80.6	yes
Qwen3.5-9B--uk--embedding--leace	Qwen3.5-9B	uk	embedding	leace (layer 16)	gender, military_status, religion	explicit, implicit, attr_free	31,050	29.4	74.3	yes
Qwen3.5-9B--uk--sft--adapter--uk_only	Qwen3.5-9B	uk	sft	qwen3.5-9b_uk_only	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	75.2	yes
gemma-4-12B-it--en--baseline	gemma-4-12B-it	en	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.0	84.4	yes
gemma-4-12B-it--uk--baseline	gemma-4-12B-it	uk	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.0	84.4	yes
gemma-4-E4B-it--en--baseline	gemma-4-E4B-it	en	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.0	75.8	yes
gemma-4-E4B-it--uk--baseline	gemma-4-E4B-it	uk	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.1	81.1	yes
lapa-v0.1.2-instruct--en--baseline	lapa-v0.1.2-instruct	en	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.0	49.5	yes
lapa-v0.1.2-instruct--en--sft--adapter--en_only	lapa-v0.1.2-instruct	en	sft	lapa-12b_en_only	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	80.6	yes
lapa-v0.1.2-instruct--uk--baseline	lapa-v0.1.2-instruct	uk	baseline	--	military_status, gender, religion, military_status_x_gender, military_status_x_religion	explicit, implicit, attr_free	161,550	0.2	52.7	yes
lapa-v0.1.2-instruct--uk--sft--adapter--uk_only	lapa-v0.1.2-instruct	uk	sft	lapa-12b_uk_only	military_status, gender, religion	explicit, implicit, attr_free	31,050	0.0	77.2	yes

Intended use and limitations

• Evaluation outputs, not training data. Every row is built from a held-out benchmark candidate. Fine-tuning on them contaminates the benchmark.
• The reference decision is attribute-free, not unbiased: it was produced by GPT-4o.
• Qwen3.5-9B English mitigation runs cover military status under the implicit condition only; other cells cover all three groups. See the GitHub report for why.
• Model outputs may contain biased hiring judgements; that is what is being measured.

Citation

TBD. The citation for the mitigation paper, and further references, will be added here once the paper is published.

The mitigation paper is in preparation; until then please cite the code repository above and the Djinni Recruitment Dataset:

@inproceedings{drushchak-romanyshyn-2024-introducing,
  title     = {Introducing the Djinni Recruitment Dataset: A Corpus of Anonymized CVs and Job Postings},
  author    = {Drushchak, Nazarii and Romanyshyn, Mariana},
  booktitle = {Proceedings of the Third Ukrainian Natural Language Processing Workshop (UNLP) @ LREC-COLING 2024},
  year      = {2024}
}