diggajupadhyay/sathi-1.7b-ne

Sathi (साथी) — a Nepali-first small language model

Sathi is a 1.7B assistant fine-tuned for Nepali: the everyday language, the Bikram Sambat calendar, lakh/crore counting, and the facts of Nepal. It runs on a laptop CPU or a low-end GPU, quantised to about 1.1 GB.

It exists because general models treat Nepali as a rounding error: they slide into Hindi, invent Bikram Sambat dates, and convert लाख into million.

Files

file	size	use
sathi-q4_k_m.gguf	~1.1 GB	llama.cpp, Ollama, LM Studio — the one most people want
Modelfile	tiny	ollama create sathi -f Modelfile

An f16 GGUF (~3.4 GB, for re-quantising at other sizes) can be added with python scripts/publish_hf.py --include-f16; it is left out by default because almost nobody needs it.

Use it

# Ollama
ollama create sathi -f Modelfile      # Modelfile is included in this repo
ollama run sathi "आजको मिति विक्रम सम्वतमा के हो?"

# llama.cpp
llama-cli -m sathi-q4_k_m.gguf -p "नेपालको राजधानी कहाँ हो?"

from llama_cpp import Llama

llm = Llama(model_path="sathi-q4_k_m.gguf", n_ctx=4096)
print(llm.create_chat_completion(
    messages=[{"role": "user", "content": "k cha halkhabar?"}]
)["choices"][0]["message"]["content"])

With the Sathi CLI, which adds the tools, the memory and the Nepal-specific data this model was trained to use:

hf download diggajupadhyay/sathi-1.7b-ne sathi-q4_k_m.gguf --local-dir models/
sathi --backend local -m models/sathi-q4_k_m.gguf "आजको मिति के हो?"

How it was trained

• Base: Qwen/Qwen3-1.7B (Apache-2.0), chosen for permissive licence, Devanagari coverage, tool-calling and size.
• Method: QLoRA (4-bit NF4 base, rank 32, all attention and MLP projections), 1 epoch, sequence length 1024, on a single 8 GB consumer GPU.
• Loss is masked to assistant turns only.
• Format: ChatML, with tool calls as <tool_call>{"name": ..., "arguments": ...}</tool_call>.

Data

Public Nepali instruction data from the Hub: saillab/alpaca-nepali-cleaned, Someman/alpaca-nepali, NepaliAI/Nepali-Health-QA, Chhabi/Nepali-Agriculture-QA, ashokpoudel/English-Nepali-Translation-Instruction-Dataset, iamTangsang/Nepali-to-English-Translation-Dataset — each under its own licence, listed in the training repository's DATA.md.

Generated locally, correct by construction, from the assistant's own implementations: Bikram Sambat conversions, Nepali numerals and number words, arithmetic word problems, tool-call and memory conversations, romanised→ Devanagari pairs, and examples of admitting ignorance.

Hindi text is filtered out of every source with a lexical Nepali/Hindi discriminator, because "Nepali" web corpora carry Hindi and that is how a Nepali model learns to answer with "है".

No scraped social media. Nothing from Reddit or other platforms whose terms forbid it is in this model.

Evaluation

Sixteen cases a Nepali assistant has to get right, scored at temperature 0 through the Sathi runtime (tools and memory on), with strict checking: numbers must match as whole tokens, and repeating the question back never counts as an answer.

capability	qwen3:1.7b (base)	llama3.2:3b	Sathi
language (Nepali not Hindi)	0/3	1/3	1/3
Bikram Sambat calendar	1/2	1/2	1/2
numerals (lakh/crore)	0/2	0/2	0/2
Nepal knowledge	2/6	4/6	3/6
reasoning	1/2	1/2	1/2
honesty	0/1	0/1	0/1
overall	4/16	7/16	6/16

Read that honestly: Sathi improves on the base it was trained from, and it is not yet better overall than Llama-3.2-3B, a model twice its size with no Nepali-specific training at all. This release is a first checkpoint. The stage that installs a language — continued pretraining on a large Nepali corpus — has not been run yet; only instruction tuning has. That is the next step, and it is where the real gains are.

The benchmark is small and readable by design; it catches the failures that make an assistant useless to a Nepali speaker rather than claiming broad coverage.

Limitations

• 1.7B parameters. It will be confidently wrong about specifics; check anything that matters, especially health, legal and financial questions.
• Nepali is its focus, not Nepal's other languages (Maithili, Bhojpuri, Tharu, Tamang, Newar, Magar and the rest) — those need their own data.
• Dates, arithmetic and Nepal reference facts are reliable only through the tools in the Sathi runtime; the raw model can still get them wrong.
• Trained on public internet data, so it carries the biases of that data.

Licence

Apache-2.0, following the base model. The training data sources keep their own licences; see DATA.md in the training repository before redistributing anything derived from them.