danilkho/ghostswitch-expert-evals

GhostSwitch expert evaluations

Do community LoRA adapters for Qwen3.5-9B actually beat the model they were trained on? This dataset holds the answers behind that question, from the expert-library check in GhostSwitch, a local assistant that attaches one LoRA expert to Qwen3.5 for each message.

13 adapters from Hugging Face were each asked 60 questions from a public benchmark of their field, and the bare base model was asked the same 60 questions. Only one adapter was significantly better than the base model.

Expert	Adapter	Benchmark	Base	Expert	Expert better / base better	p	Verdict
coding_expert	blackshell69/Qwen3.5-9B-distilled-coding-agentic	HumanEval+	76.7%	88.3%	8 / 1	0.039	verified
legal_contract_expert	ximilala/qwen35-9b-legal-lora	LegalBench CUAD	76.7%	83.3%	4 / 0	0.125	leaning better
privacy_dlp_expert	querypieai/qwen3.5-9b-dlp-pii-lora	Gretel PII (F1)	75.3%	78.3%	7 / 2	0.180	leaning better
software_engineering_expert	Shreyansh327/qwen3.5-9b-swegym-lora-full	HumanEval+	76.7%	81.7%	6 / 3	0.508	undecided
finance_market_expert	ZaoKing/khudi-finance-v0-3-arithmetic-candidate	FinQA	61.7%	66.7%	7 / 4	0.549	undecided
reasoning_expert	cudabenchmarktest/qwen3.5-9b-qwen3.6-reasoning-distilled	GSM8K (no thinking)	88.3%	90.0%	2 / 1	1.000	undecided
coding_delta_expert	danielcherubini/Qwen3.5-DeltaCoder-9B	HumanEval+	76.7%	78.3%	3 / 2	1.000	no better
text_to_sql_expert	whileai/text-to-sql-shop-qwen3.5-9b-r1-step75	Gretel text-to-SQL, run in SQLite	58.3%	58.3%	1 / 1	1.000	no better
medical_reasoning_expert	jamezoon/qwen3-5-9b-medmcqa-lora	MedQA (4 options)	66.7%	65.0%	1 / 2	1.000	no better
penetration_testing_expert	HRB25/qwen35-9b-pentest-v5	pentesting-eval	88.3%	86.7%	3 / 4	1.000	no better
iac_security_expert	selcuksntrk/iac-secfix-qwen35-9b-lora	Terraform S3 security MCQ	53.3%	51.7%	5 / 6	1.000	no better
entity_extraction_expert	illuminator22/openbrief-nano-qwen35-entity-extractor	CoNLL-2003 (F1)	67.4%	67.2%	17 / 14	0.720	no better
cybersecurity_expert	Muneerali199/rakshak-cwe-v2	CTI-Bench, CVE → CWE	56.7%	51.7%	3 / 6	0.508	no better (scored lower)

A "no better" result means only that the adapter did not beat the base model on this benchmark, in this setup. Several adapters were trained for a narrower task or a different prompt format than the one used here, so this is not a verdict on their quality for the job their authors built them for.

Setup

• Base model: Qwen3.5-9B, GGUF Q4_K_M, llama.cpp on an RTX 3090, served by GhostSwitch.
• Conditions: the same prompt for both sides; the only difference is the LoRA adapter and the expert's instructions. GhostSwitch's answer checks and thinking were off for both. Multiple-choice tasks ask for a final "Answer: X" line.
• Sample: 60 questions per expert, drawn with a fixed seed (0) from each benchmark's test split (the train split for pentesting-eval).
• Scoring: automatic. Code runs against the HumanEval+ tests, SQL runs in SQLite, numbers and choices are compared with the gold answer, entity and PII tasks use F1.
• Statistics: a paired two-sided sign test over the questions where the two answers scored differently. With 60 questions, only differences of roughly 10 points or more show up reliably, so "undecided" really means undecided.
• Date: October 2026. Script: scripts/expert_eval.py.

Configs

• summary (13 rows): one row per expert with the adapter repository and pinned revision, the benchmark, both accuracies, the paired counts, p-value and verdict.
• pairs (780 rows): one row per question per expert, with the base and expert scores, the extracted final answers, the full answer texts, retries and latency.

Question texts and gold answers are not included, because the source benchmarks have their own licenses. question_id is the benchmark's own id where it has one, otherwise the row index in the split, so the rows can be joined back to the source:

Task	Source dataset
humaneval	evalplus/humanevalplus
gsm8k	openai/gsm8k
finqa	dreamerdeo/finqa
medqa	GBaker/MedQA-USMLE-4-options
cuad	nguha/legalbench, CUAD tasks
cti_rcm	AI4Sec/cti-bench, cti-rcm
pentest	preemware/pentesting-eval
terraform	rootly-ai-labs/terraform-s3-security-mcq
pii	gretelai/gretel-pii-masking-en-v1
conll	tner/conll2003
sql	gretelai/synthetic_text_to_sql

The answer texts are model outputs and are released under Apache 2.0; the benchmarks and the adapters keep their own licenses.

Reproduce or extend

git clone https://github.com/hiddenway/GhostSwitch.git && cd GhostSwitch
python scripts/expert_eval.py run --quick --limit 60
python scripts/expert_eval.py report

A larger sample (around 150 questions) would settle the undecided experts; 14 more adapters in the GhostSwitch catalog (Verilog, OpenSCAD, Lean 4, CAD and others) still need a judge or tools to be graded. Results from other runs are welcome as a discussion or a pull request. All the adapters, with the base models, are in the GhostSwitch experts collection.