Instructions to use Nanthasit/sakthai-plus-1.5b-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Nanthasit/sakthai-plus-1.5b-lora with Transformers:

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

pipe = pipeline("text-generation", model="Nanthasit/sakthai-plus-1.5b-lora")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Nanthasit/sakthai-plus-1.5b-lora", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Nanthasit/sakthai-plus-1.5b-lora with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Nanthasit/sakthai-plus-1.5b-lora"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Nanthasit/sakthai-plus-1.5b-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Nanthasit/sakthai-plus-1.5b-lora

SGLang

How to use Nanthasit/sakthai-plus-1.5b-lora 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 "Nanthasit/sakthai-plus-1.5b-lora" \
    --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": "Nanthasit/sakthai-plus-1.5b-lora",
		"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 "Nanthasit/sakthai-plus-1.5b-lora" \
        --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": "Nanthasit/sakthai-plus-1.5b-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Nanthasit/sakthai-plus-1.5b-lora with Docker Model Runner:
```
docker model run hf.co/Nanthasit/sakthai-plus-1.5b-lora
```

SakThai Plus 1.5B — LoRA

General-purpose assistant adapter · Qwen2.5-1.5B-Instruct + LoRA · conversational + reasoning

Part of the SakThai Model Family Built from a shelter in Cork, Ireland, with $0 budget, no GPU, and a relentless drive to build something meaningful.

Training code and configs live in beer-sakthai/Sak-Family-Agent.

SakThai Plus 1.5B LoRA is a general-purpose assistant adapter built on Qwen/Qwen2.5-1.5B-Instruct. It is trained with SFT/TRL on the SakThai Combined v7 family dataset, optimized for natural conversation, reasoning, and follow-up quality while staying small enough to run on consumer hardware.

Model Description

Attribute	Value
Base model	Qwen/Qwen2.5-1.5B-Instruct
Parameters	1.5B
Adapter	LoRA (r=16, alpha=32, dropout=0.05)
LoRA target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
LoRA variant	rslora
Context length	32,768 tokens
Method	SFT via TRL
License	Apache 2.0
Output	Text generation

Usage

Apply LoRA adapter

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-1.5B-Instruct"
lora_id = "Nanthasit/sakthai-plus-1.5b-lora"

base_model = AutoModelForCausalLM.from_pretrained(base_id)
tokenizer = AutoTokenizer.from_pretrained(base_id)

model = PeftModel.from_pretrained(base_model, lora_id)
model = model.merge_and_unload()

inputs = tokenizer("You are a helpful assistant.", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Inference via merged artifact

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Nanthasit/sakthai-plus-1.5b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

prompt = "Plan a 3-day trip to Lisbon with a €200 budget."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Inference via Hugging Face Inference API

curl -X POST https://api-inference.huggingface.co/models/Nanthasit/sakthai-plus-1.5b \
  -H "Authorization: Bearer ***" \
  -H "Content-Type: application/json" \
  -d '{"inputs": "Your prompt here"}'

Benchmarks

Benchmark	Value	Status
Tool Calling — send_email (SakThai Bench v2)	1.0	verified
General Assistant — text generation	pending	multi-trial eval ongoing

Evidence files: .eval_results/cron-eval-2026-07-31-2.yaml.

See the leaderboard Space for live updates as scores are verified.

Evaluation

Conversation: Multi-turn dialogue, follow-up handling, context recall
Reasoning: Step-by-step problem-solving, creative answers
Instruction Following: SFT-tuned for clear, concise responses
Inference Speed: ~50-100 tokens/sec on CPU; ~500+ tokens/sec on GPU
Tool Use: Inherited merged sibling shows full-tool-call accuracy on send_email benchmark

Training Data: SakThai Combined v7 (2,003 examples — tool-use, Q&A, multi-turn conversation, reasoning tasks)

Training Details

Item	Value
Base	Qwen/Qwen2.5-1.5B-Instruct
Dataset	Nanthasit/sakthai-combined-v7
Framework	TRL + PEFT 0.19.1
Adapter type	LoRA + rslora
LoRA rank	16
LoRA alpha	32
Dropout	0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Task type	CAUSAL_LM
Dtype	bfloat16
Context	32,768 tokens

Reproduce Evaluation

python scripts/eval_tool_calling.py \
  --model_id Nanthasit/sakthai-plus-1.5b-lora \
  --base_model_id Qwen/Qwen2.5-1.5B-Instruct \
  --dataset Nanthasit/sakthai-bench-v3 \
  --task_type tool_calling \
  --trials 3

Reproduce Training / Merge

python scripts/train_lora.py \
  --base_model Qwen/Qwen2.5-1.5B-Instruct \
  --dataset Nanthasit/sakthai-combined-v7 \
  --output_dir lora-out

python scripts/merge_lora.py \
  --base_model Qwen/Qwen2.5-1.5B-Instruct \
  --adapter_dir lora-out \
  --output_dir merged

Limitations

LoRA has lower capacity than full fine-tuning — complex multi-step reasoning may benefit from a merged+quantized version
Best results when base model is loaded with appropriate prompt template
No instruction-only mode — always use base model's chat template
Tool-calling benchmark here is reproduced on the merged checkpoint, not adapter-only inference
Generated content can be inaccurate; verify factual claims independently

Citation

@misc{sakthai-plus-1.5b-lora,
  title  = {SakThai Plus 1.5B LoRA},
  author = {Beer Nanthasit and SakThai Agents},
  year   = {2026},
  url    = {https://huggingface.co/Nanthasit/sakthai-plus-1.5b-lora
}

Family & Links

Asset	Link
Collection	SakThai Model Family
Leaderboard	Benchmark Results
Training Dataset	SakThai Combined v7
Base Model	Qwen2.5-1.5B-Instruct
Merged Version	sakthai-plus-1.5b
Repository	beer-sakthai/Sak-Family-Agent