Nanthasit/sakthai-context-1.5b-merged

SakThai Context 1.5B

Fine-tuned from Qwen2.5-1.5B-Instruct on the SakThai combined dataset for tool-calling, multi-turn context, and instruction-following capabilities. Designed as the reasoning backbone for the SakThai agent (running on Hermes Agent framework).

Model Details

Property	Value
Base Model	Qwen/Qwen2.5-1.5B-Instruct
Architecture	Qwen2 (decoder-only transformer)
Hidden Size	1536
Layers	28
Attention Heads	12
Intermediate Size	8960
Vocab Size	151936
Fine-tuning Method	LoRA (r=16, α=32, dropout=0.1)
Target Modules	q_proj, k_proj, v_proj, o_proj
Training Steps	220
Training Duration	~39 minutes (4 epochs on 974 examples)
License	Apache 2.0

Training

• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Dataset: Nanthasit/sakthai-combined-v4 — 974 training + 51 test examples covering 25 canonical tool schemas
• Method: LoRA via PEFT (rank=16, alpha=32, dropout=0.1) on q/k/v/o projections
• Optimizer: AdamW, linear schedule, 220 steps

The LoRA adapter weights are available at Nanthasit/sakthai-context-1.5b-tools.

Evaluation

45/45 tests passed (100%) across 3 runs (15 tests/run).

Category	Tests	Pass Rate
Basic (greeting, identity)	6	✅ 100%
Multi-turn (name recall, context, preference)	9	✅ 100%
Instruction following	6	✅ 100%
Tool calling	6	✅ 100%
Reasoning (math, coding, explanation)	6	✅ 100%
Format adherence (JSON, markdown, arrays)	12	✅ 100%

Full eval report: eval/EVAL.md

Usage

Via Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Nanthasit/sakthai-context-1.5b-merged")
tokenizer = AutoTokenizer.from_pretrained("Nanthasit/sakthai-context-1.5b-merged")

messages = [{"role": "user", "content": "What's the capital of Japan?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Via Inference Client

from huggingface_hub import InferenceClient

client = InferenceClient("Nanthasit/sakthai-context-1.5b-merged")
response = client.chat_completion(
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=256
)
print(response.choices[0].message.content)

Merging the Adapter

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "Nanthasit/sakthai-context-1.5b-tools")
merged = model.merge_and_unload()
merged.save_pretrained("./sakthai-context-1.5b-merged")

Limitations

• Fine-tuned primarily for tool-calling and structured output; general knowledge remains at Qwen2.5-1.5B-Instruct baseline level.
• Tested on CPU — performance on GPU inference may produce slightly different output distributions.
• Best suited for agentic workflows with well-defined tool schemas. Complex multi-hop reasoning may require a larger base model.

Bias, Risks & Safety

This model is fine-tuned from Qwen2.5-1.5B-Instruct and inherits its base strengths and limitations. As a small language model (1.5B parameters), it may exhibit:

• Factual inaccuracies on niche or recent topics
• Biases present in the base model's pre-training data
• Limited performance on tasks requiring long context (>2K tokens) or deep multi-step reasoning

Deploy with appropriate guardrails for any user-facing application.

Citation

@misc{sakthai-context-1.5b,
  author = {Nanthasit},
  title = {SakThai Context 1.5B — Tool-Calling Fine-Tune},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Nanthasit/sakthai-context-1.5b-merged}}
}