Nanthasit/sakthai-combined-v7

SakThai Coder 1.5B 💻

Code + tool-calling · Qwen2.5-Coder-1.5B fine-tune · Q4_K_M GGUF for CPU

The code specialist of the SakThai family — Qwen2.5-Coder-1.5B fine-tuned for tool-calling and shipped as a CPU-friendly GGUF. Part of the House of Sak.

Model Description

SakThai Coder 1.5B is a Q4_K_M GGUF of Qwen2.5-Coder-1.5B-Instruct, fine-tuned on sakthai-combined-v6, sakthai-combined-v7, sakthai-bench-v2, and sakthai-irrelevance-supplement. It is optimized for code generation and tool-calling on CPU via llama.cpp / Ollama.

Key facts

Parameter	Value
Base model	Qwen/Qwen2.5-Coder-1.5B-Instruct
Architecture	Qwen2ForCausalLM (qwen2)
Parameters	~1.54 B
Hidden size	1,536
Layers	28
Attention heads	12 (GQA, 2 KV heads)
Intermediate size	8,960
Vocabulary	151,936
Context length	32,768 (32K)
RoPE theta	1,000,000
Fine-tune method	QLoRA → GGUF Q4_K_M
GGUF size	1.12 GB
License	Apache 2.0

How to Use

1) llama.cpp CLI

wget https://huggingface.co/Nanthasit/sakthai-coder-1.5b/resolve/main/qwen2.5-coder-1.5b-instruct-q4_k_m.gguf
./llama-cli -m qwen2.5-coder-1.5b-instruct-q4_k_m.gguf   -p "Write a Python function to merge two sorted lists:"   -n 256 --temp 0.2

2) Ollama

echo 'FROM ./qwen2.5-coder-1.5b-instruct-q4_k_m.gguf' > Modelfile
ollama create sakthai-coder -f Modelfile
ollama run sakthai-coder "Write a script that monitors CPU usage"

3) llama-cpp-python

from llama_cpp import Llama

llm = Llama(model_path="qwen2.5-coder-1.5b-instruct-q4_k_m.gguf", n_ctx=4096)
out = llm(
    "Write a Python function to merge two sorted lists:",
    max_tokens=256,
    temperature=0.2,
    echo=False,
)
print(out["choices"][0]["text"])

4) Tool-calling example (ChatML <tools>)

system
You are a coding assistant with tool-calling ability. Available tools:
<tools>
[
  {"name": "read_file", "description": "Read file contents", "parameters": {"type": "object", "properties": {"path": {"type": "string"}}, "required": ["path"]}},
  {"name": "run_test", "description": "Run a pytest file", "parameters": {"type": "object", "properties": {"file": {"type": "string"}}, "required": ["file"]}}
]
</tools>

User: Read test_sample.py, then write a function that passes the tests in it.

For tool-calling, keep function schemas inside <tools> and let the model emit tool calls before answering.

Benchmarks

The fine-tune starts from Qwen2.5-Coder-1.5B-Instruct, whose reported scores are:

Benchmark	pass@1	Notes
HumanEval	74.4%	Single-turn Python function completion
MBPP	71.2%	Multi-program synthesis from docstring
MultiPL-E (Python)	65.3%	Multi-language subset

Source: Qwen2.5-Coder-1.5B-Instruct evaluation. These are base-model reference values — the fine-tune has not been independently re-run on these benchmarks, so they serve as an honest reference ceiling.

Internal SakThai Coding Suite

The fine-tuned model was tested against an internal coding benchmark covering five tasks: algorithm, debugging, code explanation, refactoring, and data processing.

Task	Result
Algorithm (factorial)	Pass
Debugging	Pass
Code explanation (async)	Pass
Refactoring	Pass
Data processing (primes)	Pass
Overall	5/5

Internal test — run locally on CPU, single trial with llama.cpp Q4_K_M, temperature=0.1. Results captured 2026-07-25 and verified by SakThai agent. Single-trial results are indicative, not a third-party benchmark, and are marked verified: false in the model-index accordingly.

Limitations

• Single-trial benchmarks only — internal coding suite and tool-calling results are not multi-trial.
• Base-model references — HumanEval/MBPP/MultiPL-E values are base-model scores, not verified on this fine-tune.
• CPU inference — performance is CPU-bound; expect lower throughput than GPU deployments.
• Context cost — larger contexts use more RAM; 32K is the model ceiling but 4K–8K is safer on low-RAM machines.
• Tool-calling format — requires ChatML-style <tools> blocks; raw prompting will not trigger tool behavior.
• Language focus — examples and fine-tuning are EN-first.
• No serverless endpoint — this repo ships GGUF only; HF Inference API is not supported.
• Quantization — Q4_K_M is efficient but slightly lower quality than F16.
• Hallucination risk — tool schemas may be partially honored; verify tool calls before side effects.
• Zero-budget lineage — trained on free Colab compute; not representative of a formal benchmarking pipeline.

Citation

@misc{sakthai-coder-1.5b,
  author       = {Nanthasit Burankum},
  title        = {{SakThai Coder 1.5B — Qwen2.5-Coder Fine-tune for Code \& Tool Calling}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Nanthasit/sakthai-coder-1.5b
  note         = {GGUF Q4_K_M, CPU-friendly inference via llama.cpp / Ollama}
}

Base model citation:

@misc{qwen2.5-coder-1.5b,
  title        = {{Qwen2.5-Coder-1.5B-Instruct}},
  year         = {2024},
  publisher    = {Alibaba Cloud},
  howpublished = {\url{https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct
}

---

_{Updated by SakThai cron — "We are one family — and becoming more."}