Prashna-7B

Prashna (प्रश्न), Hindi for "question", is a 7B Socratic tutoring model fine-tuned from Qwen2.5-7B-Instruct. Instead of handing students the answer, it diagnoses what they misunderstand and asks guiding questions that lead them there.

It was built for Akara (Team Evolve AI), a pipeline that turns NCERT chapters into explainer videos and then checks understanding with the Feynman technique: the student explains the concept back, and the tutor asks a targeted follow-up.

Model Details

Model name	Prashna-7B
Base model	Qwen2.5-7B-Instruct
Parameters	~7.6B
Task	Socratic tutoring, misconception diagnosis, follow-up question generation
Language	English (see Limitations)
License	Apache 2.0 (inherited from the base model)
Developed by	rachitgoyell

Intended Use

• Socratic follow-up questions after a student explains a concept
• Diagnosing gaps between a student's explanation and a reference script
• Step-by-step scaffolding with hints that withhold the final answer
• Research on pedagogical alignment and answer leakage

Out of Scope

• Giving direct answers or solutions (it is designed to withhold them)
• Use as an authoritative source of facts without a grounding passage
• Unsupervised use with young students without teacher or parent oversight
• Anything high-stakes, such as grading or placement decisions

How to Use

pip install -U transformers accelerate torch

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "rachitgoyell/prashna-7b"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto")

reference_script = "Objects in free fall accelerate at the same rate regardless of mass, ignoring air resistance."

messages = [
    {
        "role": "system",
        "content": (
            "You are a Socratic tutor. Never state the final answer. "
            "Compare the student's explanation with the reference passage, identify the key "
            "misconception, and ask exactly one guiding question.\n\n"
            f"Reference passage:\n{reference_script}"
        ),
    },
    {"role": "user", "content": "I think heavier objects fall faster because gravity pulls them harder."},
]

inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=400, temperature=0.7, do_sample=True)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

Low-VRAM (4-bit, NVIDIA GPU)

from transformers import BitsAndBytesConfig

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16)
model = AutoModelForCausalLM.from_pretrained(repo, quantization_config=bnb, device_map="auto")

Recommended Prompting

• Put the reference passage in the system prompt so questions stay grounded
• State the rules explicitly: no final answers, one question per turn
• Pass the full conversation history so hints get progressively more specific
• Use a moderate temperature (0.5-0.8)

Limitations and Risks

• Domain coverage. Behavior outside the subjects in the training data is untested.
• Answer leakage. Research on LLM tutors shows leakage can rise sharply when students push for the answer. This model can still reveal solutions under pressure.
• Hallucination. Without a reference passage, it may produce incorrect facts or ask questions built on wrong premises.
• Noisy input. Speech-recognition errors in transcripts can cause wrong diagnoses.
• Not a teacher replacement. Use it as a support tool with human oversight.

References

• Qwen Team, Qwen2.5.
• From Problem-Solving to Teaching Problem-Solving: Aligning LLMs with Pedagogy using Reinforcement Learning, arXiv:2505.15607 (EMNLP 2025). Cite this only if your training followed its method; delete the line otherwise.

Citation

@misc{prashna-7b,
  author = {rachitgoyell},
  title  = {Prashna-7B: A Socratic Tutoring Model for Akara},
  year   = {2026},
  url    = {https://huggingface.co/rachitgoyell/prashna-7b}
}