TutorDesk Qwen3-4B
A LoRA fine-tuned version of Qwen/Qwen3-4B trained for Indian CBSE/NCERT tuition teachers (Classes 6–10, Math & Science).
Part of the TutorDesk AI project — submitted to the Hugging Face × Gradio "Build Small Hackathon" (Backyard AI track, June 2026).
What it does
This model is fine-tuned on three objectives, all grounded in the Indian CBSE/NCERT curriculum:
| Objective | Task | Examples |
|---|---|---|
| Question Generation | Given a class, subject, topic, and difficulty → output an exam question + answer + explanation | ~3,000 |
| Difficulty Classification | Given a question → output Easy, Medium, or Hard | ~1,000 |
| Indian-style Grading | Given a question, marking scheme, and student answer → award step marks + one-line feedback | ~600 |
Total training examples: ~4,600
Model details
| Property | Value |
|---|---|
| Base model | Qwen/Qwen3-4B |
| Parameters | 4B (≤4B → qualifies for Tiny Titan) |
| Fine-tune method | LoRA (PEFT) |
| LoRA rank | 16 |
| LoRA alpha | 32 |
| Target modules | q_proj, k_proj, v_proj, o_proj |
| Training epochs | 3 |
| Batch size | 4 × 4 grad accum = 16 effective |
| Learning rate | 2e-4 (cosine schedule, 5% warmup) |
| Max seq length | 2,048 tokens |
| Precision | fp16 |
| Hardware | Modal A10G (24 GB VRAM) |
| Adapter | Merged into base weights (no adapter files needed at inference) |
Training data
- Source: ParthKadam2003/NCERT_Dataset (MIT license, ~124k rows)
- Scope filtered to: Classes 6–10 · Mathematics & Science · CBSE/NCERT
- Grading triples: Synthesized using the base Qwen3-4B as a teacher model — each Q&A pair was expanded into 4 student-answer types (correct/no-working/partial/wrong-method) with step-mark breakdowns.
- All data formatted as ChatML (Qwen3 chat template,
enable_thinking=False).
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "naazimsnh02/tutordesk-qwen3-4b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")
def ask(system, user, max_new_tokens=512):
messages = [{"role": "system", "content": system}, {"role": "user", "content": user}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=max_new_tokens, temperature=0.7, do_sample=True)
return tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
Generate a question
response = ask(
"You are an expert Indian CBSE tutor for Classes 6-10 Math and Science.",
"Generate a Medium Short Answer question for Class 8 Science on the topic 'Cell Structure'. "
"Provide the full question, answer, and explanation.",
)
print(response)
Classify difficulty
response = ask(
"You classify the difficulty of Indian CBSE exam questions for Classes 6-10 as Easy, Medium, or Hard. "
"Reply with only the label — nothing else.",
"Classify difficulty for this Class 9 Science question:\n"
"Explain the process of photosynthesis and write the balanced chemical equation.",
)
print(response) # → Medium
Grade a student answer (Indian-style)
response = ask(
"You are an Indian CBSE teacher grading student answers. "
"Award step marks, give partial credit where deserved, and provide a one-sentence feedback comment.",
"Grade out of 3 marks.\n\n"
"Question: State Newton's Second Law of Motion.\n\n"
"Marking scheme:\n"
" Step 1: Force is proportional to rate of change of momentum (1 mark)\n"
" Step 2: F = ma stated or derived (1 mark)\n"
" Step 3: Units and direction mentioned (1 mark)\n\n"
"Student answer: Force equals mass times acceleration. More force means more acceleration.",
)
print(response) # → Marks: 2/3\nFeedback: ...
Scope & limitations
- Scope: Classes 6–10 · CBSE/NCERT · Mathematics & Science · English (primary), Hindi (partial)
- Not trained for: Humanities, language arts, competitive exams (JEE/NEET), classes below 6 or above 10
- Grading: Best on objective Math/Science answers; open-ended essay grading is not reliable
- Handwriting OCR: This model operates on text. Handwriting recognition is handled separately by MiniCPM-V in the TutorDesk pipeline
- Language: Outputs are primarily English; multilingual outputs handled by Tiny Aya in the TutorDesk pipeline
Citation
If you use this model, please cite the base model and dataset:
@misc{tutordesk-qwen3-4b,
author = {Naazim},
title = {TutorDesk Qwen3-4B: CBSE Tutor Fine-tune},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/naazimsnh02/tutordesk-qwen3-4b}
}
License
Apache 2.0 (inherited from Qwen3-4B base). Training data from ParthKadam2003/NCERT_Dataset (MIT).