codementor-v2-fullstack

CodeMentor V2 — Full-Stack (codementor-v2-fullstack)

A QLoRA-fine-tuned LoRA adapter that extends CodeMentor, a Socratic programming tutor, from 4 foundational languages to 17 full-stack technologies.

This adapter is Phase 2 of a two-phase continued fine-tuning pipeline. It is not trained from scratch — it loads the Phase 1 checkpoint (likithyadavv/codementor-7b) as its base and attaches a new, larger LoRA adapter on top.

Model Tree

Qwen/Qwen2.5-7B
  └── Qwen/Qwen2.5-Coder-7B
        └── Qwen/Qwen2.5-Coder-7B-Instruct
              └── likithyadavv/codementor-7b        (Phase 1: 4 languages)
                    └── likithyadavv/codementor-v2-fullstack   (Phase 2: 17 technologies — this adapter)

What This Model Does

CodeMentor is designed to teach, not just answer. Given a student's code snippet and a question, it:

1. Identifies the language or technology
2. Acknowledges what the student got right
3. Explains the problem conceptually — without handing over the corrected solution
4. Closes with a guiding question that pushes the student to reason through the fix themselves

Technologies Covered

Phase 1 (retained via dataset replay): Python, Java, C, C++

Phase 2 (new in this adapter): HTML/CSS, JavaScript, TypeScript, React, Next.js, Node.js, Express, FastAPI, Django, Spring Boot, SQL, MongoDB, Docker, Git, REST APIs

Training Details

Why load Phase 1 as the base instead of raw Qwen? The Phase 1 checkpoint already encodes Socratic tutoring behavior for 4 languages, so Phase 2 only needs to learn new technology-specific vocabulary and error patterns — not the tutoring format itself. This warm start is reflected in training loss falling below 1.0 within 50 steps, despite Phase 2 having 16× more training examples than Phase 1.

Dataset replay: 1,585 Phase 1 examples (~19.8% of the Phase 2 training mix) were included to prevent catastrophic forgetting of the original 4 languages. Post-training evaluation confirmed 100% retention on all Phase 1 language test cases.

Evaluation

Manual evaluation across all 17 technologies, one representative test prompt each, against 5 pass/fail criteria (technology identification, correct error diagnosis, Socratic closing question, no direct solution given, Phase 1 retention):

• Overall: 97.2% (70/72 checks passed)
• Technology identification: 100% (17/17)
• Socratic closing question present: 100% (17/17)
• Phase 1 language retention: 100% (4/4)

This is a manual smoke-test evaluation, not a held-out benchmark — see Limitations below.

Usage

from transformers import AutoModelForCausalLM, BitsAndBytesConfig, AutoTokenizer
from peft import PeftModel
import torch

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type='nf4',
    bnb_4bit_compute_dtype=torch.bfloat16,
)

base = AutoModelForCausalLM.from_pretrained(
    'likithyadavv/codementor-7b',
    quantization_config=bnb,
    device_map='auto',
)

model = PeftModel.from_pretrained(base, 'likithyadavv/codementor-v2-fullstack')
tokenizer = AutoTokenizer.from_pretrained('likithyadavv/codementor-v2-fullstack')

Prompt format:

### System:
You are CodeMentor, a patient programming tutor for Python, Java, C, C++, HTML, CSS,
JavaScript, TypeScript, React, Next.js, Node.js, Express, FastAPI, Django, Spring Boot,
SQL, MongoDB, Git, REST APIs, and Docker. Always identify the language or technology
first. Acknowledge what is correct, then guide with hints and questions -- never give
away the full solution.

### Instruction:
{instruction}

### Input:
{code_snippet}

### Response:

Limitations

• Evaluated manually on one prompt per technology — no automated benchmark (BLEU/ROUGE/BERTScore) or held-out test set yet
• Single-turn only — no conversation memory across exchanges
• Training data was authored against a fixed schema and quality-reviewed, but at hand-crafted scale (8,000 examples), which limits further growth without a more scalable data pipeline

Citation

If you use this model, please cite:

CodeMentor LLM: A QLoRA Fine-Tuned Socratic Programming Tutor
Mohammad Yusha G.N., Likith Yadav, Suhas R, Dhananjay M. Hiremath
(Guide: Prof. Sanjivani D. Tipe)
Dept. of Artificial Intelligence & Machine Learning, MVJ College of Engineering, Bengaluru
SYNERGY 2026 — IC-SIIT, MS Ramaiah University of Applied Sciences

Acknowledgements

Built using the Hugging Face ecosystem — Transformers, PEFT, TRL, BitsAndBytes, and Datasets. Compute provided via Kaggle's free-tier GPU program.