Concept Guru

Getting a JEE or NEET question right is only half the battle. Students preparing for India's most competitive exams need to understand why an answer is correct, not just that it is. Current language models are optimized for accuracy and conversational fluency, but they rarely teach. They skip intuition, elide the first-principles reasoning a teacher would never omit, and treat every response as a lookup rather than a lesson.

ConceptGuru is a 4B parameter multilingual language model trained to explain concepts the way an experienced teacher would. It covers Physics, Chemistry, Biology, and Mathematics across Foundation, JEE, and NEET syllabi, and does so in ten Indian languages so students can learn in the language they think in.

The problem with answer-first AI

Most language models, even capable ones, default to answer-first behavior. Given a problem, they compute a solution and present it. For a student who already understands the concept, that is fine. For a student who does not, it teaches nothing.

Effective pedagogy requires a different ordering. You build intuition before reaching for the formula. You connect the new concept to what the student already knows. You surface and correct the misconceptions that predictably arise. You illustrate with real examples and analogies. And you explain the reasoning, not just the result.

General-purpose models are rarely trained on data that rewards this behavior. ConceptGuru is.

Building a teacher-centric dataset

Teacher-quality explanations do not exist at scale in structured form. Textbooks give definitions. Internet forums give shortcuts. What we needed was the kind of explanation that happens in a classroom: layered, intuitive, deliberately paced.

We built a multilingual data generation pipeline that transforms educational lecture content into supervised training examples. The pipeline has six stages.

Stage 1: Lecture transcription

Educational lectures across Physics, Chemistry, Biology, and Mathematics are converted into structured transcripts. These are not raw captions. Each transcript is processed to preserve the pedagogically relevant signals that a teacher naturally embeds: intuitive explanations, worked examples, analogies, and explicit misconception corrections.

Stage 2: Agentic concept segmentation

Transcripts are not chunked by length. An agentic segmentation framework identifies semantically coherent concept boundaries, grouping content by instructional unit rather than by token count. This preserves the internal logic of each explanation and prevents concepts from being split at inconvenient boundaries.

Stage 3: Question generation

For each concept segment, a diverse set of questions is generated across five educational categories:

Question type	Purpose
Conceptual	Test understanding of definitions and principles
Reasoning-based	Require multi-step inference from first principles
Application	Map concepts onto novel or real-world scenarios
Exam-oriented	Mirror the structure and demands of JEE and NEET papers
Misconception-focused	Probe and correct predictable student errors

To prevent the model from overfitting to a narrow prompt distribution, questions are varied across length, format (short-answer, long-answer, MCQ-style framing), verbosity level, and tone (formal, conversational, exam-style).

Stage 4: Knowledge grounding

Each question is paired with supporting content retrieved from two sources: NCERT textbooks (curriculum-aligned, authoritative) and the original lecture transcript segments (classroom-authentic, pedagogically grounded). This RAG-style retrieval anchors the model to both the official syllabus and real teaching practice.

Stage 5: Teacher-style answer generation

A dedicated generation system synthesizes answers designed to mimic an experienced educator rather than an information retrieval system. Generated responses are evaluated on whether they lead with conceptual understanding rather than the formula, show step-by-step reasoning, use first-principles derivations where appropriate, ground claims in real-world examples, and flag examination relevance explicitly.

Stage 6: Multilingual expansion and quality assurance

All generated content is translated into ten Indian languages using frontier language models. Before human review, every sample passes through automated quality gates.

Format correctness checks markdown structure and LaTeX equation rendering. Language detection ensures target language match. An LLM-as-judge module scores instruction-following quality. Samples below threshold are regenerated or discarded. Post-filtering, subject matter experts manually evaluate translations for educational accuracy, terminology consistency, and residual errors not caught by automated checks.

Dataset statistics

Metric	Value
Subjects	Physics, Chemistry, Biology, Mathematics
Unique questions	400,000+
Samples per language	150,000+
Total multilingual samples	1.5 million+
Knowledge sources	NCERT + educational lectures

Model training

ConceptGuru is fine-tuned on Qwen3-4B-Instruct using supervised learning on the full 1.5 million sample multilingual corpus. A configurable system prompt adjusts response language at inference time.

Training optimizes for instruction following, educational reasoning, multilingual understanding, and teacher-style explanation generation. Particular weight is given to the last objective. A model that produces a one-line answer to a JEE Physics question may be correct, but it is rarely useful to a student who did not already know how to solve it.

Evaluation

ConceptGuru is evaluated across multiple dimensions against its base model, Qwen3-4B-Instruct.

Metric	Qwen3-4B-Instruct	ConceptGuru (4B)
Win rate (%)	21.4	78.6
Coverage (%)	74.2	81.6
Accuracy / factualness (%)	86.5	98.3
Multilingual consistency (%)	32.8	78.4

Win rate is evaluated via pairwise comparison where human evaluators select the preferred response based on explanation quality, clarity, and pedagogical effectiveness. Coverage measures how comprehensively a response addresses all relevant sub-topics for a given concept. Accuracy validates correctness of formulas, definitions, derivations, and reasoning against NCERT content.

The multilingual consistency result is worth noting separately. The base model scores 32.8%, meaning roughly two-thirds of its non-English responses degrade in quality compared to English. ConceptGuru brings that to 78.4%, a +45.6 percentage point improvement, which is the direct result of training on SME-validated multilingual data rather than relying on the base model's incidental multilingual ability.

What this demonstrates

A 4B model trained on teacher-derived explanations outperforms much larger general-purpose models on educational tasks. That gap is not about parameter count. It is about training data. The same model architecture, trained on the same number of steps, produces qualitatively different outputs depending on whether its training examples came from generic internet text or from structured pedagogical content.

The pipeline also shows how domain-specific data generation can capture tacit expertise. Teacher knowledge about how to explain Newton's third law or the behaviour of ionic compounds in solution is not written down in any textbook. It lives in lectures, in the specific analogies a teacher chooses, in the misconceptions they preemptively address. Extracting that knowledge and converting it into training signal is an engineering problem, and one this pipeline is designed to solve.

ConceptGuru is trained on Qwen3-4B-Instruct so it can run cost-effectively at scale. The goal is a system where every student in India can get a clear, conceptually rigorous explanation in their own language, regardless of which classroom they sit in.