Llama-3.2-3B-RecruitSheriff

Llama-3.2-3B-RecruitSheriff

Built with Llama.

A QLoRA adapter for Llama 3.2 3B Instruct that scores a resume against a job description and returns a structured verdict: a match score out of 100, strengths grounded in the resume, and gaps relative to the job.

The point of the project was to put the output contract in the weights instead of in a prompt, so scoring behaves the same way on every call and no third-party LLM API is needed at request time.

What it produces

Match Score: 60/100

Strengths:
- ...

Gaps:
- ...

Top Interview Questions:
- ...

How to prompt it

The model was trained on one exact instruction. Matching it matters: a reworded instruction caused a complete format failure during development.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "unsloth/llama-3.2-3b-instruct-unsloth-bnb-4bit"
ADAPTER = "SomkeX/Llama-3.2-3B-RecruitSheriff"

tokenizer = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER)

SYSTEM_MESSAGE = "You are an expert HR recruiter and ATS system."
INSTRUCTION = "Analyze this resume against the job description."

messages = [
    {"role": "system", "content": SYSTEM_MESSAGE},
    {"role": "user", "content": f"{INSTRUCTION}\n\nResume: {resume_text}\n\nJob Description: {jd_text}"},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=600, do_sample=False)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Greedy decoding (do_sample=False) is what the reference backend uses.

Training

No preference optimisation (DPO) was applied. This is SFT only.

Training data

600 resume and job-description pairs, generated locally with Ollama running Llama 3.2 3B from 20 role templates. No paid API was used at any stage.

The teacher and the student are the same 3B model, so data quality is bounded by what that model can judge. This is the main ceiling on the adapter's quality.

Evaluation

15 held-out resume and job-description pairs, not seen in training.

On the strengths metric, the automated checker uses word overlap, so most flagged cases were checker limitations ("leadership" not matching "led a team"). One was a real error: the model reported a job requirement as a resume strength. That case is kept in the number rather than removed.

Two bugs were found in the evaluation code itself during this work: an instruction that had drifted from the training prompt, and a grounding checker that could not read negation. Both were fixed before these numbers were published.

Baseline comparison

Against Gemini on the same 15 pairs, this adapter tracks closely on clear mismatches but is more conservative on strong-fit candidates (roughly 60 to 80 where Gemini gives 92 to 95), and its scores cluster on a coarse grid. It learned the scoring tendencies of its training data.

Limitations

• Requires a GPU. Loading without quantisation on CPU crashes rather than running slowly.
• Scores one resume against one job description per call. No batching, no accounts, no storage.
• Conservative and coarse on strong-fit candidates, as above.
• The reference FastAPI app currently fails to parse the "Top Interview Questions" section, so that field comes back empty in the app even when the model produces it. Known bug in the parser, not in the weights.
• Not a hiring decision system. It gives one structured, repeatable opinion and does not replace human judgement.

Licence

Derived from Llama 3.2 and governed by the Llama 3.2 Community License. Built with Llama.