Instructions to use henryhale/Kintu-2.0-30M-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use henryhale/Kintu-2.0-30M-Instruct with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="henryhale/Kintu-2.0-30M-Instruct")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("henryhale/Kintu-2.0-30M-Instruct")
model = AutoModelForCausalLM.from_pretrained("henryhale/Kintu-2.0-30M-Instruct", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use henryhale/Kintu-2.0-30M-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "henryhale/Kintu-2.0-30M-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "henryhale/Kintu-2.0-30M-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/henryhale/Kintu-2.0-30M-Instruct

SGLang

How to use henryhale/Kintu-2.0-30M-Instruct with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "henryhale/Kintu-2.0-30M-Instruct" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "henryhale/Kintu-2.0-30M-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "henryhale/Kintu-2.0-30M-Instruct" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "henryhale/Kintu-2.0-30M-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use henryhale/Kintu-2.0-30M-Instruct with Docker Model Runner:
```
docker model run hf.co/henryhale/Kintu-2.0-30M-Instruct
```

Kintu-2.0-30M-Instruct

A 30M-parameter Luganda chat model: Kintu-2.0-30M, further fine-tuned on multi-turn Luganda chats and then preference-tuned with DPO. The architecture and tokenizer are unchanged.

Training

SFT: luganda-chats and luganda-lima, with long chats split into windows of whole exchanges that fit 512 tokens, each keeping its system message: 2,327 training sequences. 5 epochs, 32 sequences per step, learning rate 5e-5 with cosine decay, fp16, on 1× T4 in 3 minutes. Training loss went from 2.64 to about 1.19.
DPO: luganda-ultrafeedback-dpo, keeping only pairs where the chosen reply scores at least 1 point higher, neither reply contains code, and both fit 512 tokens: 13,659 of 44,496 pairs. 1 epoch, β = 0.1, learning rate 1e-6, 32 pairs per step, fp16, on 1× T4 in 10 minutes. DPO loss went from 0.69 to about 0.65.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("henryhale/Kintu-2.0-30M-Instruct")
model = AutoModelForCausalLM.from_pretrained("henryhale/Kintu-2.0-30M-Instruct")

messages = [
    {"role": "system", "content": "Oli muyambi ow'amagezi ayamba abantu mu Luganda."},
    {"role": "user", "content": "Oli otya?"},
]
enc = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True)
out = model.generate(**enc, max_new_tokens=256, do_sample=True, temperature=0.4, top_p=0.8, repetition_penalty=1.2)
print(tok.decode(out[0, enc["input_ids"].shape[1]:], skip_special_tokens=True))

Limitations

At 30M parameters it often drifts off-topic and gets facts wrong, so don't rely on it for information, especially on health, money or farming. The fine-tuning data is mostly AI-written and machine-translated, and hasn't been fully reviewed by native speakers. Keep chats short: it saw at most 512 tokens at a time.