रिसर्च पेपर, हिंदी में
Learn machine-learning papers in Hindi.
Watch a short Hindi video on the paper and another on the tool, check yourself with a quick quiz — then build something with what you learned.
- 1Watch & readA Hindi video, then the paper with a guide.
- 2ConceptsThe few ideas you should be able to explain.
- 3Quick quizTwo minutes to check it really landed.
- 4ActionA small hands-on task with a real result.
Track 1 · 9 steps
Fine-tune Qwen for GST
Take a small open Qwen model and make it measurably better at reasoning about Indian GST than it is out of the box — on a free Colab GPU, with the data, the model and the scores published on Hugging Face.
Papers in this track
-
LoRA
Fine-tune by adding a small low-rank update, not rewriting the model.
-
QLoRA Hindi video coming soon
Freeze the base in 4-bit, train adapters on top — a fine-tune on the GPU you already have.
-
LIMA Hindi video coming soon
A thousand carefully chosen examples can be enough.
-
Textbooks Are All You Need Hindi video coming soon
Better data beats a bigger model.
-
Self-Instruct Hindi video coming soon
Bootstrap thousands of training examples from a handful of seeds — and inherit the generator's mistakes if you do not filter.
-
Quality, Diversity, Complexity Hindi video coming soon
A survey of what actually makes synthetic data useful — three properties that pull against each other.
-
Evol-Instruct (WizardLM) Hindi video coming soon
Make a question harder on purpose — add a constraint, a reasoning step, a rarer case.
-
RAFT Hindi video coming soon
Train the model with the right document mixed in among distractors, so it learns to pick.
-
Agentic Adversarial QA 2026 Hindi video coming soon
Generate the questions where a weak model and a strong model disagree — a small set aimed at interpretive reasoning.
-
QueST Hindi video coming soon
Train the question generator itself to produce difficult problems.
-
InstructGPT Hindi video coming soon
The three-step template every assistant follows — demonstrations, preferences, RL.
-
Tulu 3 Hindi video coming soon
A full post-training recipe, published — every stage, every dataset.
-
DPO Hindi video coming soon
Learn from "this answer is better than that one" — without a reward model.
-
Distillation Hindi video coming soon
Teaching a small model to think like a big one.
-
Who Judges the Judge? 2026 Hindi video coming soon
An LLM judge's verdict shifts with which model wrote the answer — validate it against an expert first.
-
DeepSeek-R1 Hindi video coming soon
Reasoning trained by reinforcement — the model is rewarded for results, not shown answers.
-
Rubric rewards with GRPO 2026 Hindi video coming soon
Reinforcement fine-tuning of a quantized Qwen with LoRA, scored criterion by criterion against a rubric.
-
Rethinking Rubric Generation 2026 Hindi video coming soon
A vague rubric gives a vague reward — how to split criteria until they actually separate good from bad.
-
Small Models as Rubric Judges 2026 Hindi video coming soon
The judge runs thousands of times during RL — a small one can be good enough and far cheaper.
-
Judge-R1 2026 Hindi video coming soon
A legal system that pairs agentic retrieval of statutes with rubric-guided RL to cut invented citations.
Hands-on lessons in this track
-
Build a small eval and an LLM judge Hindi video coming soon
Turn your failure types into test questions, and check a judge against an expert before you trust its scores.
huggingface.co docs 1 ↗docs.claude.com docs 2 ↗ Practise · step 1
-
Your first Unsloth notebook on Colab Hindi video coming soon
Run a complete LoRA fine-tune end to end and know what each cell is doing.
-
Pick a Qwen size and load it in 4-bit Hindi video coming soon
Choose the largest model your GPU can train, and measure its memory instead of guessing.
-
Chat templates — the format the model expects Hindi video coming soon
Render a conversation into the exact text and special tokens Qwen was trained on.
unsloth.ai docs 1 ↗huggingface.co docs 2 ↗unsloth.ai docs 3 ↗ Practise · step 4
-
Generate a dataset and push it to the Hub Hindi video coming soon
Go from a few seed examples to a few hundred generated ones, inspect them, and publish the set.
huggingface.co docs 1 ↗distilabel.argilla.io docs 2 ↗ Practise · step 5
-
The training run — settings and the loss curve Hindi video coming soon
Set the handful of numbers that matter, read the loss curve, and fix the usual failures.
unsloth.ai docs 1 ↗huggingface.co docs 2 ↗unsloth.ai docs 3 ↗ Practise · step 7
-
Train on preference pairs, then run the model on your laptop Hindi video coming soon
Turn expert corrections into DPO training data, and export the result to GGUF.
Track 2 · 9 steps
Faster RAG with Opus and Sonnet
Build a retrieval-augmented GST assistant, measure it honestly, then make it faster and cheaper without making it worse — by sending each question to the smallest Claude model that can handle it and not paying twice for the same tokens.
Papers in this track
-
RAFT Hindi video coming soon
Train the model with the right document mixed in among distractors, so it learns to pick.
-
Judge-R1 2026 Hindi video coming soon
A legal system that pairs agentic retrieval of statutes with rubric-guided RL to cut invented citations.
-
RAG Hindi video coming soon
The original — look up passages first, then generate from them.
-
Dense Passage Retrieval Hindi video coming soon
Find passages by meaning, not shared words — the idea behind every embedding search.
-
RAGAS Hindi video coming soon
Score a RAG system in parts — was the right text found, and did the answer stick to it?
-
How Much Do Legal RAG Systems Still Hallucinate? 2026 Hindi video coming soon
Eight legal RAG systems measured claim by claim — and questions with a false premise are where they break.
-
RouteLLM Hindi video coming soon
A small router decides which model answers each question.
-
Hybrid LLM Hindi video coming soon
Predict how hard the query is; send only the hard ones to the big model.
-
FrugalGPT Hindi video coming soon
Run the cheap model first — then check its work.
-
AutoMix Hindi video coming soon
Let the small model grade its own answer against the retrieved context.
-
RouterBench Hindi video coming soon
One ruler for every router — and a humbling result.
-
Temporal Misgrounding in Legal RAG 2026 Hindi video coming soon
Tax-law RAG keeps citing today's version of a provision when the question is about another date.
-
The Token Tax 2026 Hindi video coming soon
Retrieve a few passages, or load every document into the prompt? An accuracy-versus-cost comparison.
-
LLMs for Austrian VAT Law Hindi video coming soon
Fine-tuning and RAG compared head to head on real VAT cases from a tax firm — the closest study to this project.
Hands-on lessons in this track
-
Build a baseline RAG pipeline Hindi video coming soon
Chunk the documents, embed them, retrieve, and answer with Claude — the plain version you will later speed up.
huggingface.co docs 1 ↗anthropic.com docs 2 ↗ Practise · step 1
-
Evaluate RAG in two halves Hindi video coming soon
Measure whether the right passage was found, and separately whether the answer stuck to it.
-
Prompt caching and batch processing with Claude Hindi video coming soon
Stop paying, and waiting, for the same long prompt again and again.
docs.claude.com docs 1 ↗docs.claude.com docs 2 ↗ Practise · step 7
This site is an experiment in learning in public — its author is the first student. Want to go deeper? Browse all illustrated papers, the illustrated decoder, or the blog.