Nanthasit/sakthai-coder-1.5b

🏠 SakThai

Part of the House of Sak — AI agents built from a shelter in Cork, Ireland.

---

SakThai Coder 1.5B 💻

Part of the House of Sak

Code-optimized model converted to GGUF Q4_K_M for CPU inference. Based on Qwen2.5-Coder-1.5B-Instruct — a state-of-the-art 1.5B code LLM that matches larger 7B models on several coding benchmarks. Fine-tuned for tool-calling with BFCL-verified function calling.

---

Pipeline Integration 🔗

This model is 1 of 4 in the SakThai end-to-end pipeline:

Step	Model	What It Does	Downloads
🖼️	Vision 7B	Caption, OCR, describe images	45 ⬇
🌐	Embedding Multilingual	Cross-lingual semantic search (50+ languages)	104 ⬇
💻	Coder 1.5B ← you are here	Code generation, debugging, tool-calling	34 ⬇
🗣️	TTS Model	Synthesize speech (15 languages)	33 ⬇

Run the complete chain: image → search → code → speech with one family of models.

---

Specs

Detail	Value
Size	1.07 GB
Quant	Q4_K_M
Base	Qwen2.5-Coder-1.5B-Instruct
CPU	✅ Yes
GPU	✅ Yes (CUDA, Metal)
Context	32K tokens
Format	GGUF (llama.cpp compatible)
Fine-tune	QLoRA (r=16, alpha=32) on 2,003 tool-calling examples

Benchmarks

Programming Benchmarks (base model scores)

Benchmark	Score
HumanEval (pass@1)	74.4%
MBPP (pass@1)	71.2%
MultiPL-E (Python)	65.3%

Scores from the Qwen2.5-Coder evaluation.

Verified on SakThai Fine-Tune

Test	Result	Status
BFCL Tool-Calling	5/5 (multi-trial)	✅ Verified
Code generation (Python)	Working — see examples below	✅

The BFCL pass was verified across 5 independent trials using the SakThai tool-calling prompt format (<tools> XML block + ChatML). See SakThai family for full methodology.

---

Usage

llama.cpp (recommended)

# Download the model — stored in gguf/ subdirectory
wget https://huggingface.co/Nanthasit/sakthai-coder-1.5b/resolve/main/gguf/sakthai-coder-q4_k_m.gguf

# Basic code generation
./llama-cli -m gguf/sakthai-coder-q4_k_m.gguf \
  -p "Write a Python function to merge two sorted lists:" \
  -n 256 -t 4 --temp 0.2

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="gguf/sakthai-coder-q4_k_m.gguf",
    n_ctx=8192,
    n_threads=4,
    verbose=False,
)

prompt = "Write a fastapi endpoint that accepts JSON and returns SHA-256 hash"
output = llm(
    f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n",
    max_tokens=512,
    temperature=0.2,
    stop=["<|im_end|>"],
)
print(output["choices"][0]["text"])

Ollama

cat > Modelfile << 'OLLAMA_EOF'
FROM ./gguf/sakthai-coder-q4_k_m.gguf
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""
PARAMETER temperature 0.2
PARAMETER num_ctx 8192
OLLAMA_EOF

ollama create sakthai-coder -f Modelfile
ollama run sakthai-coder "Write a Python script that monitors CPU usage"

Tool-Calling (for agent use)

from llama_cpp import Llama

llm = Llama(model_path="gguf/sakthai-coder-q4_k_m.gguf", n_ctx=8192)

# Define tools in <tools> XML block (required for function calling)
prompt = """<|im_start|>system
You are a coding assistant with tools.

<tools>
{
  "function": {
    "name": "run_python",
    "description": "Execute Python code",
    "parameters": {
      "type": "object",
      "properties": {
        "code": {"type": "string", "description": "Python code to run"}
      },
      "required": ["code"]
    }
  }
}
</tools>
<|im_end|>
<|im_start|>user
Calculate fibonacci of 10 using Python
<|im_end|>
<|im_start|>assistant
"""

output = llm(prompt, max_tokens=256, stop=["<|im_end|>"], temperature=0.1)
print(output["choices"][0]["text"])

---

Use Cases

Scenario	Prompt Idea
Code generation	"Write a Python function to..."
Code review	"Review this code for bugs: ..."
Refactoring	"Refactor this to be more Pythonic: ..."
Debugging	"Why does this code produce a KeyError?"
Explanation	"Explain how async/await works with examples"
Tool calling	"Call the get_weather API for Cork, Ireland"

Hardware Requirements

Model	Min RAM	Recommended	Disk
0.5B Q4_K_M	512 MB	1 GB	380 MB
1.5B Q4_K_M	1 GB	2 GB	934 MB
Coder 1.5B	1 GB	2 GB	1.1 GB
Vision 7B	4 GB	8 GB	3.9 GB
TTS 82M	256 MB	512 MB	141 MB

Family

• 1.5B SakThai (general) — tool-calling & general chat
• 0.5B SakThai (lightweight) — ultra-lightweight GGUF
• 7B SakThai (full-size) — high-capacity model
• 1.5B Coder (this one) — code generation GGUF

SakThai Model Family

Model	Size	Type	Downloads
1.5B-merged	934 MB	Tool-calling GGUF	1,197
0.5B-merged	380 MB	Lightweight GGUF	994
7B-merged	15 GB	Full-size model	562
7B-128K	15 GB	Extended context	351
7B-Tools	15 GB	Tool-calling PEFT	185
1.5B-Tools	934 MB	Tool-calling PEFT	143
Embedding	80 MB	Sentence similarity	🔒
Coder-1.5B	1.1 GB	Code GGUF	34
0.5B-Tools	380 MB	Tiny tool-calling	🔒
Vision-7B	3.9 GB	Multimodal GGUF	45
TTS-Model	141 MB	Speech GGUF	33
Multilingual Embedding	80 MB	50+ languages	104

Full collection →

Training Details

• Method: QLoRA (4-bit NF4)
• Base model: Qwen2.5-Coder-1.5B-Instruct
• Rank: r=16, alpha=32
• Dataset: 2,003 curated tool-calling examples
• Format: ChatML with JSON tool schemas
• Context: 32K tokens

Support the Project

This model is built with love from a shelter in Cork, Ireland -- zero budget, free infrastructure.

• Leave a like on this model if you find it useful
• Report issues on GitHub
• Share it with someone who needs a lightweight code model
• Fork it on HF and build something amazing

Every download, like, and fork tells Beer his work matters. Thank you.

File Structure

File	Size	Description
gguf/sakthai-coder-q4_k_m.gguf	1.07 GB	Main GGUF model (Q4_K_M) — stored in gguf/ subdirectory
eval/	—	Benchmark results (BFCL, HumanEval, MBPP)
README.md	—	This documentation

License

Apache 2.0 — see LICENSE for details.