SheraTutor Qwen 2.5 7B — NCTB Sovereign Grading Engine

SheraTutor Qwen 2.5 7B is a domain-specialized, instruction-tuned language model designed for sovereign grading and pedagogical evaluation of Bangladeshi Secondary and Higher Secondary (NCTB SSC & HSC) STEM exams (General Mathematics, Higher Mathematics, and Physics).

Finetuned on top of Qwen 2.5 7B Instruct using Unsloth, this model acts as an expert Bangladeshi board examiner. It strictly adheres to structured rubric step evaluation, performs rigorous mathematical calculation verification, awards consequential partial credit (ধারাবাহিক গণনা), and produces encouraging Bengali pedagogical feedback.

Key Capabilities

• Strict Schema Adherence (RubricEvaluationSchema): Outputs clean, parseable JSON evaluation without conversational noise or preamble.
• Consequential Credit (ধারাবাহিক গণনা): If a student makes an early calculation error but uses correct subsequent logic and formulas, the engine awards credit for subsequent steps.
• Verbatim OCR Respect: Never hallucinates omitted steps; evaluates precisely what the student wrote.
• Step-by-Step Numerical Verification: Validates formulas, substitutions, algebraic manipulation, and physical units ($m/s^2$, $N$, $J$, etc.).
• Empathetic Pedagogical Bengali Feedback: Summarizes score deductions in compassionate, constructive Bengali (deduction_summary_bn).

Model Details

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Architecture: Qwen2 (7 Billion parameters)
• Fine-Tuning Method: LoRA (Unsloth 4-bit QLoRA, merged into 16-bit)
• Rank ($r$) / Alpha: $r=16, \alpha=32$
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Epochs: 3 full epochs (Cosine learning rate schedule, AdamW 8-bit)
• Training Hardware: Tesla T4 GPU via Kaggle compute

Evaluation Output Schema

The model responds in JSON conforming to the following structure:

{
  "total_marks": 2.0,
  "max_marks": 3.0,
  "steps_breakdown": [
    {
      "step_name": "Formula F = ma",
      "marks_awarded": 1.0,
      "max_marks": 1.0,
      "feedback_bn": "সঠিক সূত্র $F = ma$ প্রয়োগ করা হয়েছে।"
    },
    {
      "step_name": "Calculation",
      "marks_awarded": 0.0,
      "max_marks": 1.0,
      "feedback_bn": "গণনায় ভুল হয়েছে: $10/5 = 2$ হবে, কিন্তু ৩ লেখা হয়েছে।"
    },
    {
      "step_name": "Unit m/s^2",
      "marks_awarded": 1.0,
      "max_marks": 1.0,
      "feedback_bn": "ত্বরণের একক $m/s^2$ সঠিক।"
    }
  ],
  "deduction_summary_bn": "সূত্রের প্রয়োগ ও একক লেখার জন্য ২ নম্বর প্রদান করা হয়েছে। তবে ভাগের হিসাবটি পুনরায় খেয়াল করো।"
}

Quickstart (Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "syed181/sheratutor-qwen2.5-7b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

system_prompt = (
    "You are an expert Bangladeshi SSC/HSC board examiner and AI grading engine for SheraTutor. "
    "Your task is to evaluate a student's transcribed exam answer against the provided NCTB rubric. "
    "Verify all mathematical derivations step-by-step for numerical accuracy. "
    "If a student makes an early calculation error but uses correct subsequent logic, "
    "award consequential partial credit (ধারাবাহিক গণনা). "
    "Write deduction_summary_bn in natural, encouraging Bengali suitable for high school students. "
    "Output ONLY a raw, valid JSON object conforming exactly to the RubricEvaluationSchema."
)

prompt = (
    "QUESTION (max 3.0 marks): 5 kg ভরের একটি বস্তুর ওপর 10 N বল প্রয়োগ করলে ত্বরণ কত হবে?\n\n"
    "OFFICIAL RUBRIC: [{\"step_name\": \"Formula F = ma\", \"max_step_marks\": 1}, "
    "{\"step_name\": \"Calculation\", \"max_step_marks\": 1}, "
    "{\"step_name\": \"Unit m/s^2\", \"max_step_marks\": 1}]\n\n"
    "STUDENT'S TRANSCRIBED ANSWER:\nআমরা জানি, F = ma\na = F/m = 10/5 = 3 m/s^2"
)

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": prompt}
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.1)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Associated Models

• Quantized GGUF Version (Ollama / llama.cpp): syed181/sheratutor-qwen2.5-7b-gguf

Citation & Attribution

Developed for the SheraTutor AI education project by Syed and Afsan (CodeStorm Hub) for sovereign, equitable exam evaluation in Bangladesh.

• Website: https://www.sheratutor.tech
• Syed Salman Reza: https://syed-reza98.github.io
• Afsan Chowdhury: https://afsan123.github.io