License notice This project is source-available under the PolyForm Noncommercial License 1.0.0 — free for personal, research, teaching, and internal-evaluation use.

Any commercial use, redistribution as part of a commercial product, or paid hosted deployment requires a separate commercial license. See COMMERCIAL.md or contact https://github.com/AnubhabBanerjee.

Unauthorized commercial use or redistribution is a violation of the license terms.

Qwen is not covered by the license above. The checkpoint this tutorial loads, Qwen2.5-Coder-1.5B-Instruct, is released by the Qwen team under the Apache License 2.0. That Apache 2.0 license is the license that applies to the Qwen model weights. See the Qwen license file.

---

---

How to make your own JEV model

A small classification head on an open Qwen model. It does not write a paragraph. It reads a Python function once and returns match or mismatch, plus a confidence.

This is the companion repo for a beginner tutorial. The new head is a few thousand numbers. The Qwen backbone stays frozen. Training the head takes a few minutes on one GPU.

The twist

add can multiply, and the file is still valid Python. An editor can see an unused note. It has no check for a name that does not mean what the body does.

check_names.py only reads functions written at the top of the file. A method on a class, or a function nested inside another function, is left alone.

The point

A chatbot guesses the next word until a sentence exists, and then you have to find the decision inside it. This model keeps Qwen's understanding and throws away the word-guessing head. One pass. One label.

Quickstart

Requires Python 3.10+.

pip install -r requirements.txt

python train.py                  # freeze Qwen, train the new head
python check_names.py sample_module.py
python quantize.py               # optional: pack the backbone to int8 on CPU

Run train.py before check_names.py. The checker reads artifacts/jev_head.pt. That file does not exist until training finishes.

The first run downloads Qwen/Qwen2.5-Coder-1.5B-Instruct into the normal Hugging Face cache on your machine. If your weights already live in a folder of your own, set CACHE_DIR at the top of model.py to that folder. Leave it as None to use the default cache.

check_names.py exits with status 1 when at least one top-level name does not agree with its body. Status 0 means every top-level function was a match. sample_module.py is expected to exit 1. Ten of its twenty-seven functions are flagged.

Layout

File	What it does
model.py	Loads the Qwen backbone and attaches the two-score head
train.py	Trains only that head on data/names.csv
check_names.py	Scores the top-level functions in a Python file
quantize.py	Packs the backbone's linear layers to int8 and prints the size
sample_module.py	The demo file, with honest names and dishonest ones
data/names.csv	Labeled functions. The same name appears with both a right body and a wrong one

Where the numbers come from

train.py prints how many numbers were allowed to move. Expect a few thousand trainable parameters, against the billion-plus in Qwen that stayed frozen.

check_names.py sample_module.py flags 10 of 27 functions. add, is_even, and maximum are mismatches. rectangle_area is a match, even though it contains an unused note an editor will complain about.

quantize.py prints the backbone size before and after packing. On a CPU float32 copy, one run went from 6.17 GB to 0.93 GB, and add still came back mismatch. Name checking in this repo still uses the ordinary backbone, so the table above stays easy to reproduce.