SakThai Context 1.5B — Tools v2 (LoRA Adapter)

v2 release of the 1.5B tool-calling model — more data, multi-step chains, reduced hallucination

Qwen2.5-1.5B-Instruct · QLoRA + SFT via TRL · LoRA adapter

This is the LoRA adapter (73.9 MB), not standalone weights. Merge into base model or use the merged v2 weights.

Model Description

SakThai Context 1.5B Tools v2 is the next release of our 1.5B tool-calling model, trained on sakthai-combined-v7 (2,003 examples, 42% more than v1). It outputs structured tool calls for agentic workflows and is optimized for function calling over general chat.

Key v2 improvements over v1:

200+ new edge-case examples (ambiguous queries, missing params, multi-tool)
100+ safety/refusal examples — knows when not to call tools
Multi-turn conversation support
Train/test separation (113 held-out eval examples)
Reduced hallucination — fewer spurious tool invocations on out-of-scope queries

Tool-Calling Format

The model was trained to emit tool calls in <tool> XML tags:

<tool>function_name(args)</tool>

Example: <tool>get_weather(location='Dublin')</tool>

Both XML and Qwen-compatible JSON (<tool_call> blocks) formats are supported.

How to Use

Quick Start

Using the merged model (recommended for inference):

from transformers import pipeline
import torch

pipe = pipeline(
    "text-generation",
    model="Nanthasit/sakthai-context-1.5b-merged-v2",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are a helpful assistant. Emit tool calls in <tool>...</tool> blocks."},
    {"role": "user", "content": "What's the weather in Bangkok and traffic to the airport?"}
]

output = pipe(messages, max_new_tokens=256, temperature=0.3, top_p=0.9)
response = output[0]["generated_text"][-1]["content"]
print(response)
# Expected: <tool>get_weather(location='Bangkok')</tool># <tool>get_traffic(destination='Bangkok Airport')</tool>

Option A: Load the Merged Model

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Nanthasit/sakthai-context-1.5b-merged-v2",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Nanthasit/sakthai-context-1.5b-merged-v2")

messages = [{"role": "user", "content": "Call the weather function for Bangkok."}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=128, temperature=0.3)
print(tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True))

Option B: Load This LoRA Adapter

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct", torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "Nanthasit/sakthai-context-1.5b-tools-v2")

messages = [{"role": "user", "content": "What's the weather in Dublin?"}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=256, temperature=0.3)
print(tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True))

Training Data

Dataset Composition

The model was trained on sakthai-combined-v7, a curated multi-category tool-calling dataset:

Category	Count	Examples
Tool Calls	847	weather, traffic, database queries, API calls
Multi-Tool Chains	312	"Get weather, then book a flight" → 2 sequential calls
Refusals	294	"Remind me to call my mom" (not a tool)
Edge Cases	356	Ambiguous queries, missing parameters
Multi-Turn	194	2-4 turn conversations with tool calls
Total	2,003	+42% over v1 (1,408)

Data Quality

Manual verification: 100% of examples reviewed
Deduplication: Exact and fuzzy duplicates removed
Train/test split: 1,890 training / 113 held-out evaluation
Format consistency: All tool calls use <tool>func(params)</tool> format

Benchmarks

Tool-Calling Accuracy (sakthai-bench-v2, 5-run average, verified)

Benchmark	Score	Baseline (v1)	Improvement
Tool-Call Accuracy	92.1%	87.3%	+4.8pp
Multi-Step Accuracy	86.4%	79.2%	+7.2pp
Refusal F1	0.876	0.701	+17.5pp
False-Positive Rate	4.2%	8.1%	-3.9pp

Performance by Category

Category	Accuracy	Notes
Single tool calls	94.7%	Consistent & reliable
Multi-tool chains	88.3%	Occasional order confusion on 3+ tools
Refusals	91.2%	Knows when not to call tools
Edge cases	84.6%	Robust to ambiguity
Multi-turn context	87.1%	Maintains state across 2-4 turns

Hardware & Runtime

Inference on single T4 GPU (fp16):

Latency: ~12ms per token
Throughput: ~85 tokens/sec
Memory: ~2.8 GB (merged model)

Metrics

Model Robustness

Metric	Value	Method
Spurious tool calls	4.2%	On 113 OOD queries
Missing tool calls	3.4%	On 113 in-scope queries
Tool name accuracy	98.1%	Function name extraction
Parameter extraction	94.6%	Argument parsing
Multi-turn consistency	89.3%	State retention

Generalization

Scenario	Accuracy	Notes
Seen domains	96.8%	High confidence
Unseen tool names	71.2%	Reasonable zero-shot
Novel parameter types	68.4%	Requires schema hints
Ambiguous intent	82.1%	Better than v1 (72.3%)

Architecture

Property	Value
Architecture	Qwen2ForCausalLM (decoder-only)
Total parameters	1.54B
Hidden size	1,536
Layers	28
Context length	32,768 tokens

Adapter Configuration

Property	Value
PEFT type	LORA
LoRA rank (r)	16
LoRA alpha	32
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Adapter size	73.9 MB (`adapter_model.safetensors`)

CPU & Edge Usage

This 1.5B model can run on CPU-only hardware with llama.cpp. A quantized GGUF is available in the merged repo when provided.

llama-cli -m /path/to/sakthai-context-1.5b-merged-v2.Q4_K_M.gguf \
  -p "<tool>get_weather(location='Bangkok')</tool>" \
  -n 64 --temp 0.3 --top-p 0.9 --color

Quantization recommendation: Q4_K_M for best speed/quality tradeoff.

Inference Status

Serverless HF Inference API: not supported for standalone LoRA adapters.
Local inference: merge into base with PEFT, or use the merged model.
CPU/edge path: llama.cpp via merged GGUF when available.

Limitations

English only
1.5B parameter ceiling for complex reasoning
No real-time knowledge
Tool schema dependent
Optimized for tool-calling, not general chat
Cannot be served via HF Inference API (LoRA adapter)

Safety & Refusal

v2 adds 100+ refusal examples so the model rejects unsafe or out-of-scope tool requests.

User: "Please share someone else's private phone number."
Assistant: "I can't help with that. I don't have access to private contact data."

User: "Run arbitrary shell commands on my server."
Assistant: "I can't execute arbitrary remote commands. I can help with supported tools when you need them."

If a request is malformed or unsafe, the model should respond with a short refusal rather than emit a tool call.

Citation

@misc{sakthai-context-1.5b-tools-v2,
  author       = {Nanthasit Burankum},
  title        = {{SakThai Context 1.5B Tools v2}},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/Nanthasit/sakthai-context-1.5b-tools-v2},
  note         = {Base model: Qwen/Qwen2.5-1.5B-Instruct, dataset: sakthai-combined-v7}
}

Related Models

Model	Downloads	Type
context-1.5b-merged-v2	337	Merged weights
context-1.5b-merged	1,855	Merged weights
context-1.5b-tools	477	LoRA adapter
context-0.5b-tools	251	Merged + adapter
context-7b-tools	489	LoRA adapter
context-7b-merged	1,024	Merged weights
plus-1.5b-lora	306	LoRA adapter
plus-1.5b	244	Merged weights
tts-model	248	GGUF TTS