SakThai 7B — Tools (LoRA)

Highest-capability tool-calling adapter · Qwen2.5-7B · for the strongest function calling

The PEFT LoRA adapter that adds tool-calling to Qwen2.5-7B — the family's strongest function-calling variant. Provides the highest accuracy on complex multi-tool workflows. Part of the House of Sak.

What it is

LoRA weights (r=16, alpha=32) for Qwen2.5-7B-Instruct, trained on sakthai-combined-v6 + sakthai-combined-v7 + irrelevance-supplement. For a single ready-to-run GGUF checkpoint, use the merged model: sakthai-context-7b-merged.

Quick start

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

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct", torch_dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "Nanthasit/sakthai-context-7b-tools")
merged = model.merge_and_unload()  # optional: bake in for inference

Training


Base model	Qwen/Qwen2.5-7B-Instruct
Method	QLoRA (4-bit)
LoRA config	r=16, alpha=32, targets q/k/v/o_proj
Data	combined-v6 (1,378) + combined-v7 (2,309) + irrelevance-supplement (60)
Format / context	ChatML with tool schema - 32K tokens

SakThai Model Family

Models

Model	Size	Role	Downloads
context-1.5b-merged	934 MB	Flagship tool-calling GGUF	1,269
context-0.5b-merged	380 MB	Lightweight / edge GGUF	1,030
context-7b-merged	15 GB	Full-power reasoning	585
context-7b-128k	15 GB	128K long-context	382
context-7b-tools (this model)	LoRA	7B tool-calling adapter	219
embedding-multilingual	80 MB	Cross-lingual embeddings	188
context-1.5b-tools	LoRA	1.5B tool-calling adapter	163
vision-7b	3.9 GB	Image-to-text (LLaVA)	104
coder-1.5b	1.1 GB	Code generation	70
tts-model	141 MB	Text-to-speech, 15 langs	69
sakthai-embedding	80 MB	English embeddings (private)	34
context-0.5b-tools	LoRA	Edge tool-calling v1	7
context-0.5b-exp-lora-masked-v4	LoRA	Experimental V4 LoRA (0.5B)	0

Growing the Ecosystem

These assets have <100 downloads but are production-ready. Every download validates quality:

Asset	Downloads	Type	Why try it
sakthai-irrelevance-supplement	0	Dataset	60 examples teaching models to decline out-of-scope calls
sakthai-combined-v7	0	Dataset	2,309 examples, 86 tool schemas, safety coverage
sakthai-bench-v1	0	Dataset	Balanced BFCL-style benchmark (235 rows, 4 categories)
sakthai-bench-v2	0	Dataset	Multi-domain benchmark (500 rows, 5 categories)
context-0.5b-tools (v1)	7	Model	Original edge tool-calling -- runs on Raspberry Pi
sakthai-embedding	34	Model	English-only semantic search, 80 MB footprint
tts-model	69	Model	Multi-language TTS, 15 languages
context-0.5b-exp-lora-masked-v4	0	Model	Experimental V4 LoRA training run

Datasets

Dataset	Downloads	Role
sakthai-combined-v6	175	Primary tool-calling training data (1,378 examples)
sakthai-kaggle-notebooks	103	ML learning notebooks for the Food-Penguin project
SimpleToolCalling	52	Deprecated -- kept for reproducibility
food-penguin-v1	51	FineWeb-Edu food classifier training data
sakthai-combined-v7	0	Expanded tool-calling (2,309 ex, 86 tools, safer)
sakthai-irrelevance-supplement	0	Teaches models to decline out-of-scope tool calls
sakthai-bench-v1	0	Balanced BFCL-style benchmark (235 rows, 4 categories)
sakthai-bench-v2	0	Multi-domain benchmark (500 rows, 5 categories)

Spaces

Space	Role
sakthai-vision-demo	Vision demo -- try the multimodal model live
sakthai-tts	TTS web demo (static)
sakthai-leaderboard	Family benchmark leaderboard (static)

12 models - 8 datasets - 3 Spaces -- full collection

Links

House of Sak - GitHub - All models - All datasets

License

Apache 2.0 (following the Qwen2.5 base model license).

Evaluation

Not independently benchmarked. Earlier versions of this card carried a model-index score derived from a small internal spot check (typically 5 or 8 hand-picked examples) presented as a benchmark result. Those entries have been removed rather than left to propagate through Hub metadata.

For tool-calling models in this family, the benchmark to use is sakthai-bench-v2 — 500 rows, balanced across simple / parallel / irrelevance, with held-out tools and multi-turn coverage. Results will be published here once this model has been run against it.