Nanthasit/sakthai-plus-1.5b

SakThai Plus 1.5B

House of Sak — enhanced tool-calling instruct model, fine-tuned from Qwen2.5-1.5B-Instruct with rsLoRA and merged to full weight.

Model Description

SakThai Plus 1.5B is a rsLoRA-merged instruction-tuned model for tool calling and agentic workflows. It builds on the SakThai Context lineage with improved tool-use precision, conversational grounding, and instruction adherence.

• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Fine-tuning: Full-parameter merge of rsLoRA adapters trained on curated tool-calling datasets
• Architecture: Qwen2ForCausalLM — 28 layers, 12 attention heads, 2 KV heads, hidden size 1536, intermediate 8960
• Parameters: ~1.54B (bf16, 2.88 GiB on disk)
• Format: Qwen2.5 chat template (standard)
• Context length: 32K tokens (max_position_embeddings: 32768)
• License: Apache 2.0

Intended Use

• Tool/function calling agents and assistants
• Conversational AI with structured output generation
• Lightweight on-device agent deployments (CPU-friendly at 1.5B)
• Prototyping agentic workflows before scaling to larger models

Training Data

The model was fine-tuned on a curated mixture of:

Dataset	Rows	Description
Nanthasit/sakthai-combined-v7	2,309	Multi-source tool-calling examples (subset of v10)
Nanthasit/SimpleToolCalling	2,002	Structured function-calling examples

Training method: rsLoRA (rank-stabilized) via TRL SFTTrainer, merged to full weights.

Hyperparameter	Value
r / alpha / dropout	16 / 32 / 0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj (all 7 linear)
Optimizer	AdamW (8-bit)
Learning rate	2e-4
Epochs	3
Precision	bf16
Hardware	Free T4 GPU (Kaggle / Colab / HF Inference), $0 budget

Performance

Benchmark results across tool-calling, reasoning, and math tasks:

Task	Metric	Score	Verified	Source
Tool Calling (send_email)	Success Rate	1.0 (3/3)	✅ Yes	llama.cpp 3-trial
Tool Calling	Valid JSON Args	1.0 (3/3)	✅ Yes	llama.cpp 3-trial
Tool Calling	Correct Answer	1.0 (3/3)	✅ Yes	llama.cpp 3-trial
Tool Selection (v2)	Selection Accuracy	84.8%	❌ No	sakthai-bench-v2
Tool Selection (v2)	Arguments Accuracy	33.7%	❌ No	sakthai-bench-v2
Tool Selection (v2)	Strict Accuracy	33.7%	❌ No	sakthai-bench-v2
WinoGrande (WSC)	Accuracy	59.6%	❌ No	lighteval
HellaSwag	Accuracy	34.0%	❌ No	lighteval
GSM8K (Math)	Accuracy	50.9%	❌ No	lighteval

Key takeaways:

• Tool calling performs reliably — 3/3 on every trial in the verified send_email scenario. The model correctly formats <tool_call> XML with valid JSON arguments.
• Tool selection is strong (84.8% selection accuracy on sakthai-bench-v2) but argument precision needs refinement (33.7% strict).
• Commonsense and math scores reflect the base Qwen2.5-1.5B-Instruct capabilities with minor degradation from tool-calling specialization.

⚠ Unverified scores are single-trial results from .eval_results/. Multi-trial replication is planned. The 3-trial tool-calling benchmark is the only verified result.

Quick Start

Python — Direct Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

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

messages = [
    {"role": "system", "content": "You are a helpful assistant with tool-calling capabilities."},
    {"role": "user", "content": "What's the weather in Bangkok?"}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.3, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Expected output format:

<tool_call>{'name': 'get_weather', 'arguments': {'location': 'Bangkok'}}</tool_call>

Python — Tool-Calling with Multiple Tools

This model uses the <tool_call> XML dialect (v2 format). Provide tool definitions in the system prompt:

messages = [
    {"role": "system", "content": """You have access to these tools:
- get_weather(location: str) — get weather for a city
- send_email(to: str, subject: str, body: str) — send an email
- search_web(query: str) — search the internet

Respond with the appropriate tool call in <tool_call> XML format."""},
    {"role": "user", "content": "Send an email to Beer with the subject 'Status update' and body 'The model is running well.'"}
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.3, do_sample=True)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
# Extracts: <tool_call>{'name': 'send_email', 'arguments': {...}}</tool_call>

llama.cpp / Ollama (GGUF)

For CPU inference, use the llama.cpp compatible GGUF version or convert manually:

# Convert to GGUF via llama.cpp
git clone https://github.com/ggml-ai/llama.cpp
cd llama.cpp
python3 convert_hf_to_gguf.py Nanthasit/sakthai-plus-1.5b --outfile sakthai-plus-1.5b-f16.gguf

# Run with llama-cli
./llama-cli -m sakthai-plus-1.5b-f16.gguf \
  -p "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nWhat's the weather in Bangkok?<|im_end|>\n<|im_start|>assistant\n" \
  -no-cnv --simple-io -n 256 -t 2

Model Family

This model is part of the SakThai Model Family — 25 models across sizes and specializations:

Variant	Type	Description
sakthai-plus-1.5b	✅ This	Merged instruct, tool calling
sakthai-plus-1.5b-lora	LoRA adapter	Lightweight adapter for the same base
sakthai-plus-1.5b-coder	Skeleton	Code-specialized variant (base: Qwen2.5-Coder-1.5B)
sakthai-context-1.5b-tools-v2	Model	Improved tool-use variant
sakthai-context-1.5b-merged-v2	Model	Full weights, merged from v2 LoRA
Family size	25 public models across merged, LoRA, vision, embedding, TTS, and coder variants
Current downloads	244
Last updated	2026-07-31
View full collection →

Limitations

• Language: Optimized for English tool-calling scenarios. Multilingual prompts may degrade performance.
• Knowledge cutoff: The base Qwen2.5-1.5B-Instruct has a knowledge cutoff of ~2024. The fine-tuning dataset (2025–2026 tool-calling examples) does not expand world knowledge.
• Argument precision: sakthai-bench-v2 shows 84.8% selection accuracy but only 33.7% strict accuracy — the model often selects the right tool but generates malformed arguments.
• Hallucination risk: Like all language models, it may generate plausible-sounding but incorrect tool arguments. Always validate tool calls before execution in production.
• Not a general reasoning model: Benchmarks on commonsense (HellaSwag 34.0%, WinoGrande 59.6%) are below larger models. This is a tool-calling specialist, not a general-purpose reasoner.
• Unverified benchmarks: 8 of 9 metric entries are single-trial and unverified. Verified scores (3/3 tool-calling) were measured on llama.cpp q4_k_m (CPU, 2 threads).
• Free-tier constraints: Trained entirely on free T4 GPU credits. No production-grade training infra was used.

Citation

@misc{sakthai-plus-1.5b,
  author = {Burankum, Nanthasit},
  title = {SakThai Plus 1.5B: Enhanced Tool-Calling Language Model},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Nanthasit/sakthai-plus-1.5b
  note = {Fine-tuned from Qwen2.5-1.5B-Instruct with rsLoRA}
}

License

Apache 2.0