Upload README.md

README.md

@@ -1,16 +1,27 @@
 ---
-title: Scratchq 110
-emoji: 💬
+title: ScratchQ 110M
+emoji: 🧠
 colorFrom: yellow
-colorTo: purple
+colorTo: green
 sdk: gradio
-sdk_version: 6.5.1
+sdk_version: 4.44.0
 app_file: app.py
 pinned: false
-hf_oauth: true
-hf_oauth_scopes:
-- inference-api
-short_description: Indian quiz trained
+license: mit
 ---
 
-An example chatbot using [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/v0.22.2/en/index), and the [Hugging Face Inference API](https://huggingface.co/docs/api-inference/index).
+# ScratchQ — 110M Indian-GK chatbot, trained from scratch
+
+A 110M-parameter Llama-style language model **trained entirely from scratch** on $2 of compute (2B FineWeb-Edu tokens), then SFT'd on 6,686 Indian general-knowledge MCQs and SQuAD passages. RAG over a 4,600-chunk Indian Wikipedia knowledge base.
+
+## What's interesting
+
+- **Trained from scratch**: random init → tokenizer → pretraining (10,500 steps, loss 3.10) → SFT (loss 1.81) — no foundation model anywhere in the pipeline except the Llama-2 *tokenizer*.
+- **Beats GPT-2 medium (3× larger)** on Indian-MCQ benchmarks: 36.5% vs 17.5% log-likelihood accuracy.
+- **Specialized scope** — politely declines questions outside Indian-context (history, mythology, geography, politics, sports, food).
+
+## Limits (be charitable, it's 110M params)
+
+- Can be wrong on niche facts even with retrieval.
+- Free-form chat accuracy on factoid questions: ~30-40% — meaningful signal but not a real chatbot.
+- Code/repo: https://github.com/rsumit123/private-llm