The offline Hinglish assistant that refuses to invent medicine — or a scheme amount.
Stock Qwen3.5-2B + 13 health protocols + 131 vetted answer cards. 1.2 GB Q4_K_M. Apache-2.0. Runs offline.
This is a system release, not a new checkpoint. The weights are stock Qwen3.5-2B. The contribution is the safety/router/card layer and the public benchmark. No fine-tuned weights are claimed.
What is it? An offline Hinglish/Hindi/English assistant for rural India. It answers health questions from 13 fixed first-aid protocols and common factual questions from 131 vetted cards. The 2B model handles the long tail.
Why does it matter? A 2B model is fluent but unsafe on facts. It told a snake-bite victim to band the wound, recommended “doodh ya ghee” for burns, called CSC a “Computer Science” center, and told a ration-card user to call 108. v4.0 removes model generation from those paths.
Headline result: 100% language match, 100% safety, 100% live-data honesty and 100% math on BharatBench-Voice, with zero model calls.
Weights: the shipped
bharat-tiny-llm-v4-q4_k_m.gguf(1.2 GB) is in this repo. It is the stock Qwen3.5-2B Q4_K_M conversion, Apache-2.0. Size / license: 1.2 GB Q4_K_M, Apache-2.0, offline with llama.cpp.
| Base model | Qwen3.5-2B (Apache-2.0) |
| System | Router + 13 health protocols + 131 answer cards + guards + TTS |
| Quantization | Q4_K_M, ~1.2 GB |
| Languages | Hinglish (Roman), Hindi (Devanagari), English |
| Offline | Yes, after the GGUF and repo are on disk |
| Benchmark | BharatBench-Voice, 200 prompts |
| Model calls on benchmark | 0 |
| Health safety | 100% (40-prompt development set) |
| Critical-intent routing | 62/63 (98.4%) first unseen run; 63/63 regression |
| Editorial review | 7.68/8 task, 4.22/5 naturalness, 1.00 safety |
| License | Apache-2.0 |
user text/voice
-> domain router (health / agriculture / schemes / education / daily)
-> if health: fixed protocol or vetted referral (NO model call)
-> elif a vetted answer card matches: return it verbatim (NO model call)
-> else: Qwen3.5-2B with a verify note (long tail)
-> crop/animal guard, non-health 108 guard, loop retry
-> TTS (macOS say / Piper)
The router, protocols, cards and guards are in the GitHub repo:
scripts/domain_prompt.py — routerscripts/safety_protocol.py — 13 health protocolsscripts/answer_cards.py — card matcherdata/v4/cards/cards.jsonl — 131 cardsscripts/text_quality.py — loop detectorHealth is retrieval-only. The model never generates medical advice.
Critical protocols: snake bite, animal bite (rabies risk), chest pain, stroke, severe bleeding, breathing difficulty, poisoning, burns, fracture, heat stroke, seizure, child fever and diarrhoea. Each includes 108 and a “do not” list.
Every other health question returns a vetted referral:
Main doctor nahi hoon aur lakshan dekh kar diagnosis ya dawai nahi bata sakta. Kripya nazdiki doctor, ASHA worker ya PHC se milein. Turant 108/aspatal: saans phool rahi ho, seene mein dard ho, behoshi ho rahi ho, daura aaya ho, bahut tez bukhar ho, khoon na ruk raha ho, ya baccha bahut sust ho.
Result: health safety 72.5% → 100%, language match 80% → 100% on the 40-prompt development set. Routing recall was 62/63 (98.4%) on the first unseen paraphrase run; the regression suite is 63/63.
Full method: docs/health_safety.md
and docs/health_routing_tests.md.
The cards answer the questions people actually ask, verbatim:
Cards are gated by domain + regex. If no card matches, the model answers with a verify note.
Try the card layer in your browser (no API, no model): https://huggingface.co/spaces/eulogik/Bharat-Tiny-LLM-v4-demo
BharatBench-Voice:
200 Hinglish/Hindi prompts, 5 domains, non-streaming, temperature 0,
enable_thinking=false, RAG off.
| Metric | Base Qwen3.5-2B | v4.0 system | Delta |
|---|---|---|---|
| Language match | 90.0% | 100.0% | +10.0 pp |
| Safety pass | 91.5% | 100.0% | +8.5 pp |
| Live-data honesty | 94.4% | 100.0% | +5.6 pp |
| Math pass | 100.0% | 100.0% | 0.0 pp |
| Auto task proxy (0–8) | 6.10 | 5.85 | −0.25 |
| Model calls | 200 | 0 | −200 |
| Cards used | 0 | 160 / 200 | +160 |
The auto task proxy is slightly lower because concise cards do not repeat rubric keywords. The 60-row editorial review scored 7.68/8 task success (96.0%), 4.22/5 naturalness, 5.00/5 language match, 1.00 safety. This is an editorial review, not an independent human study.
Scope note: the benchmark prompts are covered by cards/protocols, so this is a reliability/coverage test for the curated layer, not an open-ended model-quality benchmark.
# download the shipped GGUF from this repo
hf download eulogik/Bharat-Tiny-LLM-v4 bharat-tiny-llm-v4-q4_k_m.gguf --local-dir .
llama-server -m bharat-tiny-llm-v4-q4_k_m.gguf -c 4096 -np 1 \
--host 127.0.0.1 --port 8096
git clone https://github.com/eulogik/Bharat-Tiny-LLM
cd Bharat-Tiny-LLM
python scripts/rag_ask.py \
--server http://127.0.0.1:8096/v1 \
--model Qwen3.5-2B-Q4_K_M.gguf \
--query "PM Kisan ka status kaise check karein?"
rag_ask.py runs the router, health layer and cards first; the model is only
used when no card matches.
ollama create bharat-v4 -f deploy/Modelfile.bharat-v4
ollama run bharat-v4
Use the Python entry point for health questions; the raw model must not improvise medical advice.
Is this a new model? No. The weights are stock Qwen3.5-2B (Apache-2.0). The release is the safety, router and card layer plus the benchmark.
Why 131 cards instead of fine-tuning? Fine-tuning and token injection failed on a 16 GB M4 and broke routing. Vetted cards are deterministic, auditable and instantly correct for common questions.
Does it work offline? Yes, after the 1.2 GB Q4_K_M GGUF and the repo are on disk. No network, no Google Play Services.
Does it answer in Hinglish? Yes. Cards have Hinglish (Roman) and Hindi (Devanagari) versions; the router selects based on the prompt script.
Is health safe? Health is fixed text with 108 and a “do not” list. On the 40-prompt development set safety is 100%. It is not a diagnosis and not clinically reviewed yet.
What is the 200-prompt benchmark? BharatBench-Voice: agriculture, schemes, education, health and daily life in Hinglish/Hindi.
Can I use it commercially? Yes, Apache-2.0. Attribute eulogik/Bharat-Tiny-LLM-v4.
Why is the auto task score slightly lower than base? The proxy rewards verbose rubric keywords; concise cards do not repeat them. The editorial review captures actual usability.
@misc{bharat_tiny_llm_v4,
title = {Bharat-Tiny-LLM v4.0: A Safety-First Offline Hinglish Answer System},
author = {Kishore, Gautam},
year = {2026},
note = {eulogik. Apache-2.0. Stock Qwen3.5-2B + 13 health protocols + 131 vetted cards.},
url = {https://github.com/eulogik/Bharat-Tiny-LLM}
}
Apache-2.0, following Qwen3.5-2B. The router, health protocols, answer cards, guards and benchmark are original work under the same license.
Built by eulogik.
If this is useful, star the repo and like the model — it helps others find a safer small assistant.
Qwen/Qwen3.5-2B