Aaa123aaa123aaa/leadership-traits-classifier

Candidate Evaluation Model

A fine-tuned Mistral-7B model specialized for evaluating university admission essays and video interview transcripts. Built for the admissions committee.

Unlike general-purpose sentiment analyzers, this model identifies high-value behavioral patterns and distinguishes authentic personal experiences from generic or AI-generated responses.

---

Model Description

Property	Value
Base model	unsloth/mistral-7b-instruct-v0.3-bnb-4bit
Fine-tuning method	LoRA (via Unsloth)
Quantization	Q4_K_M (GGUF)
Training data	726 labeled interview answers (364 strong / 362 weak)
Training time	~2 hours on RTX 4060 8GB
Language	English only (IELTS/TOEFL required for inVision U)
Output format	JSON with scores and explanations

---

Evaluation Framework

The model scores candidates across three weighted pillars:

1. Motivation — 35%

Identifies the depth of the candidate's connection to the program and their purpose.

Target signals:

• Personal triggers — "I witnessed", "because I saw", "that specific moment"
• Program alignment — references to specific labs, alumni, or tracks
• Community-centric goals — "my region", "back home", "my village"

Negative indicators:

• Vague prestige language — "excellent reputation", "prestigious institution"
• Generic motivation applicable to any university
• Cliché openings — "from a young age", "I have always been passionate"

---

2. Leadership — 35%

Evaluates the "give more than take" principle — proactive social contribution without expectation of reward.

Target signals:

• Quantifiable achievements — "team of 12", "8 months", "40% reduction"
• Unsolicited initiative — "no one asked me", "on my own", "without being asked"
• Altruistic patterns — "for free", "without pay", "gave back", "not for credit"
• Concrete results — "9 out of 12 passed", "wait times dropped by half"

Negative indicators:

• Abstract leadership definitions without evidence — "I have always been a leader"
• Self-proclaimed titles — "natural leader", "born leader"
• No concrete examples or numbers

---

3. Creativity — 30%

Assesses the uniqueness of the candidate's perspective and narrative voice.

Target signals:

• Non-standard response structures
• Unexpected angles of approach — "what most people", "instead of", "rather than"
• High-impact personal details that make the response memorable

Negative indicators:

• Predictable academic structure — "Firstly... Secondly... In conclusion..."
• Over-reliance on transition words — "Furthermore", "Moreover", "It is worth noting"
• Response could have been written by any applicant

---

Authenticity & AI Detection

Three parallel metrics run alongside the scoring:

Metric	What it measures
Burstiness	Sentence length variability — human speech is unpredictable, AI text is monotonic
Type-Token Ratio (TTR)	Lexical diversity — authentic writing uses richer vocabulary
GPT Marker Analysis	Scans for 30+ linguistic markers common in LLM-generated text

---

Output Format

The model returns a structured JSON object:

{
  "motivation_score": 88,
  "leadership_score": 92,
  "creativity_score": 75,
  "overall_score": 85,
  "key_evidence": [
    "founded tutoring program",
    "12 students",
    "8 months without pay"
  ],
  "red_flags": [],
  "committee_note": "Strong response. Candidate demonstrates unsolicited initiative with quantified outcomes. Clear give-more-than-take leadership pattern. Overall: 85/100."
}

---

Example Inputs

High-scoring input (90+)

"When I was 17, the water pump in our village broke and no one came to fix it for three months. I started mapping wells in a 10km radius using Google Maps and a bicycle. I helped 12 families. No one asked me to. When I showed the map to the local council, they actually used it. That's when I understood that data can change decisions. inVision U's systems design track is the only program I found that teaches exactly this."

Why it scores high: specific age, concrete problem, measurable impact (12 families, 10km), unsolicited initiative, explicit alignment with inVision U's specific program.

---

Low-scoring input (20–35)

"Throughout my academic journey, I have always been passionate about making a positive impact. From a young age, I recognized the importance of education. Furthermore, my leadership experiences have taught me collaboration. In conclusion, I am committed to bringing my unique perspective to your prestigious institution."

Why it scores low: no specific details, generic prestige language, 6+ GPT markers, structure applicable to any university.

---

Training Details

Base model  : unsloth/mistral-7b-instruct-v0.3-bnb-4bit
Method      : LoRA fine-tuning (Unsloth)
LoRA rank   : 16
LoRA alpha  : 16
Epochs      : 3
Batch size  : 2 (effective: 16 with gradient accumulation)
Learning rate: 2e-4
Scheduler   : cosine
Dataset     : 726 examples (364 strong / 362 weak)
Output      : GGUF Q4_K_M (~7GB)

The training dataset consists of interview answers labeled as strong or weak based on the inVision U selection rubric. Strong answers contain concrete stories with numbers, names, and real actions. Weak answers contain generic phrases that could apply to any institution.

---

Limitations & Ethical Disclosure

• Decision support only. This model is intended to highlight high-potential candidates — not as a sole instrument for automatic rejection. All final decisions remain with the admissions committee.
• English only. The model is optimized for English-language inputs. inVision U requires IELTS or TOEFL, so all interviews are conducted in English.
• Human-in-the-loop required. Every AI recommendation can be overridden by committee members. All overrides are logged to an audit trail.
• No PII stored. Candidate names, emails, and demographic data are excluded from model inputs. Only anonymized IDs are used.
• Bias guard. A separate BiasGuard module runs before and after scoring to detect and flag potential demographic bias in outputs.

---

Citation

Built with Mistral-7B + Unsloth LoRA + llama-cpp-python