Nanthasit

sft-out

README

License: apache-2.0

Model Description

Table with columns: Property, Value
Property	Value
Base model	Nanthasit/sakthai-context-0.5b-tools (Qwen2.5-0.5B-Instruct)
Adapter type	LoRA
LoRA rank `r`	8
LoRA alpha	16
Dropout	0.0
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`
Adapter size	~158 KB
Max context	32,768 tokens
Parameters	495M base + adapter
Language	English

This adapter is a low-rank delta over the base SakThai 0.5B tool-calling model. It preserves the same architecture and tokenizer.

What's in the box

Table with columns: File, Size, Description
File	Size	Description
`adapter_model.safetensors`	138 KB	LoRA delta weights
`adapter_config.json`	1.1 KB	Adapter hyperparameters
`training_args.bin`	129 B	Serialized training args
`tokenizer.json`	13.2 KB	Tokenizer

Total adapter size: ~158 KB

How to Use

Merge with base model

python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base = "Nanthasit/sakthai-context-0.5b-tools"
adapter = "Nanthasit/sft-out"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

Generate:

python
messages = [{"role": "user", "content": "What's the weather in Tokyo?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=128, temperature=0.3, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Merge to GGUF (llama.cpp)

bash
python -m peft.utils.merge_adapters \
  --base_model Nanthasit/sakthai-context-0.5b-tools \
  --adapter Nanthasit/sft-out \
  --output merged-model
python -m transformers.convert_save_to_gguf merged-model --outtype q4_k_m --outfile sakthai-0.5b-sft-out-q4_k_m.gguf

Then run with llama.cpp:

bash
./llama-cli -m sakthai-0.5b-sft-out-q4_k_m.gguf -p "What's the weather in Tokyo?" -n 128 --temp 0.3

Tool-Calling Format

This adapter was trained on the SakThai tool-calling format.

Prompt format:

xml
<tools>
{
  "function": "get_weather",
  "params": {"location": "Bangkok"}
}
</tools>

Model output:

json
{"name": "get_weather", "arguments": {"location": "Bangkok"}}

Multi-turn tool results:

xml
<tool_result>
{"result": "The weather in Bangkok is 32°C and sunny."}
</tool_result>

Benchmarks

No published benchmarks for this specific adapter yet. The base model Nanthasit/sakthai-context-0.5b-tools achieves strong single-shot tool-calling on sakthai-bench-v2.

For the SakThai family benchmarks, see the SakThai Leaderboard.

Reproduce

bash
pip install transformers>=5.14.1 trl>=1.9.2 peft datasets accelerate
python -m trl.cli.sft \
  --model_name_or_path Nanthasit/sakthai-context-0.5b-tools \
  --dataset_name Nanthasit/sakthai-combined-v7 \
  --adapter_output_dir sft-out \
  --per_device_train_batch_size 8 \
  --learning_rate 2e-4 \
  --num_train_epochs 3 \
  --lora_r 8 --lora_alpha 16 --lora_dropout 0.0 \
  --lora_target_modules q_proj,k_proj,v_proj,o_proj \
  --bf16

Training Details

Table with columns: Attribute, Value
Attribute	Value
Base model	Nanthasit/sakthai-context-0.5b-tools
Method	Supervised Fine-Tuning (SFT) via TRL
TRL version	1.9.2
Transformers	5.14.1
PyTorch	2.11.0
Datasets	5.0.0
Tokenizers	0.22.2

Note: Exact training hyperparameters (learning rate, batch size, epochs) are not available from this repo alone. The training_args.bin file is present but contains unparsed binary data. Check sibling model sakthai-context-0.5b-tools for comparable settings.

Limitations

Adapter-only. Must be loaded on top of the base model. Cannot be used standalone via AutoModelForCausalLM.
No standalone GGUF. Must be merged with base before GGUF conversion.
Not servable serverless. The Hugging Face Inference API does not support dynamic LoRA adapter loading.
Experimental. This is a new SFT training experiment; performance has not been extensively evaluated.
Placeholder tokenizer. The tokenizer.json in this repo is a placeholder; the actual tokenizer lives in the base model repo.
License compatibility depends on base-model license (apache-2.0).

Working Alternatives

Table with columns: Model, Type, Best for
Model	Type	Best for
sakthai-context-0.5b-tools	Base adapter (8.7 MB)	Strongest standalone 0.5B tool caller
sakthai-context-0.5b-merged	Merged (988 MB)	Direct inference, no merge needed
sakthai-context-0.5b-tools-sft-v2	Adapter (8.3 MB)	Newer SFT variant with richer data
sakthai-context-1.5b-merged

Citation

If you use this adapter or the SakThai family in your work, please cite:

bibtex
@misc{sakthai-sft-out,
  title  = {SakThai SFT-out LoRA Adapter},
  author = {Beer (beer-sakthai) and the SakThai Agent family},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Nanthasit/sft-out}}

Base model:

bibtex
@misc{qwen25-2025,
  title  = {Qwen2.5: A Party of Foundation Models},
  author = {Qwen Team},
  year   = {2025},
  url    = {https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct}
}

TRL:

bibtex
@software{vonwerra2020trl,
  title   = {{TRL: Transformers Reinforcement Learning}},
  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = {2020}
}

SakThai Family

This adapter is part of the SakThai model family.

Table with columns: Model, Downloads, Type
Model	Downloads	Type
context-1.5b-merged	1,855	Merged weights
context-0.5b-merged	1,692	Merged weights
context-7b-merged	1,024	Merged weights
context-0.5b-tools

Downloads verified 2026-07-31.

Built with zero budget. Part of the House of Sak.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.