qa-engineer-v1

qa-engineer-v1

QLoRA fine-tune of unsloth/qwen2.5-3b-instruct-unsloth-bnb-4bit, trained to write software test plans like a 6-7 year experienced QA engineer. This checkpoint is merged (LoRA folded into the base weights) -- loadable directly with transformers, no PEFT or Unsloth required at inference time. See requirements.txt in this repo for the minimal inference dependencies.

Intended use

Prompt the model with this fixed instruction followed by a software requirement:

You are a senior QA engineer with 6-7 years of professional testing experience. Given the following software requirement, produce a comprehensive, well-structured test plan. Your response MUST use exactly these Markdown section headings, in this order: ## Requirement Summary & Assumptions, ## Missing Requirement Suggestions, ## Risk Analysis, ## Functional Test Cases, ## Negative Test Cases, ## Boundary Value Test Cases, ## Edge Cases, ## Security Test Cases, ## API Test Cases, ## Performance Test Cases, ## Accessibility Test Cases, ## Cross-Browser / Device Test Cases, ## Test Data Suggestions. For the API Test Cases section, write 'Not applicable — no API surface described in this requirement.' if the requirement is purely UI-only. Each test case must be written as: **TC-<PREFIX><NN>: <Title>** followed by a short bulleted list with Preconditions, Steps, Expected Result, and Priority (High/Medium/Low). Use ID prefixes: F for functional, N for negative, B for boundary, E for edge, S for security, A for API, P for performance, AC for accessibility, X for cross-browser/device.

Output is a fixed 13-section Markdown test plan (Requirement Summary & Assumptions, Missing Requirement Suggestions, Risk Analysis, Functional/Negative/Boundary/Edge/Security/API/Performance/Accessibility/Cross-Browser Test Cases, Test Data Suggestions).

Training

• Method: QLoRA via Unsloth, LoRA r=16, alpha=16, target_modules=[o_proj, k_proj, q_proj, up_proj, gate_proj, v_proj, down_proj]
• Dataset: data/versions/v1/train.jsonl (60 examples)
• Epochs: 3.0
• Final training loss: 1.490
• Trained fresh from the base model (not continual/chained across versions).
• Never trained on the held-out benchmark set used for evaluation below.

Evaluation

Scored on 25 held-out benchmark items, blending rule-based checks with a local LLM judge (8 rubric categories, 0-10 each).

• Hallucination rate: 0.48
• Missed edge case count: 22/25
• Missed security case count: 23/25

License

Inherits the base model's license -- check unsloth/qwen2.5-3b-instruct-unsloth-bnb-4bit before any commercial use.