SomkeX

RecruitSheriff-Qwen3-4B

README

License: apache-2.0

Files

Table with columns: File, Use
File	Use
`recruitsheriff-lora-f16.gguf`	adapter for llama.cpp (66 MB), used by the app
`adapter_model.safetensors`, `adapter_config.json`	the same adapter for PEFT / transformers
`loss_curve.png`	training and validation loss

Run with llama.cpp (base GGUF from unsloth/Qwen3-4B-Instruct-2507-GGUF):

text

llama-server -m Qwen3-4B-Instruct-2507-Q4_K_M.gguf --lora recruitsheriff-lora-f16.gguf -c 8192

The app sends a JSON schema with each request, so the answer is always valid JSON.

Training

Three tasks: job description -> requirement checklist, resume -> verdicts with quotes, verdicts -> interview questions.
Data: 100 Indian job descriptions and 780 synthetic resumes (from strong matches to mismatches, plus tricky ones) graded by a larger teacher model (DeepSeek V4 Flash): 1,128 examples, 1,063 for training and 65 for validation. No real resumes.
QLoRA with Unsloth: r=16, alpha=32, 2 epochs, lr 2e-4 cosine, batch 1 x 8, max length 4096, loss on answers only. 88 minutes on one RTX 4060 laptop GPU (8 GB). Validation loss 0.483 -> 0.430 and never went up.

Results

Held-out test set: 10 job descriptions the model never saw, 58 synthetic resumes, labels from the teacher and then reviewed by hand.

Table with columns: System, Verdicts correct, Fake quotes, Score error (points), Fit label correct, Ranking (Spearman)
System	Verdicts correct	Fake quotes	Score error (points)	Fit label correct	Ranking (Spearman)
This adapter	92%	2%	3.1	97%	0.93
Qwen3-4B base, same prompt	80%	6%	9.4	78%	0.83
Teacher (upper bound)	99%

Bias check: swapping only the candidate's name (gender or community) moved scores by -0.3 and -0.1 points on average, so no direction; single borderline resumes moved by up to 11 points, and 1 fit label flipped.
5 real resumes (shared with consent, never used for training) x 2 real job descriptions: 74% of verdicts agreed with labels made from the resume text alone (base model: 68%), 0 fake-quote claims (base: 6). The labels came from a second AI reviewer, not a recruiter, and the sample is tiny.

Limits

Made for English resumes and Indian job descriptions; a student project, not a hiring decision tool.
Weak on requirements with several parts ("communication, teamwork and leadership"), and sometimes marks a skill "met" when it only appears in a skills list.
A human should read the quotes before deciding anything about a candidate.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

SomkeX

Model Tree

Base

Qwen/Qwen3-4B-Instruct-2507

Adapter

this model

Input Modalities

Text

Output Modalities