README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - sakthai
  - house-of-sak
  - notebooks
  - fine-tuning
  - kaggle
datasets:
  - Nanthasit/sakthai-combined-v6
  - Nanthasit/sakthai-combined-v7
  - Nanthasit/sakthai-bench-v1

SakThai Kaggle Notebooks & Deploy Scripts

The exact training notebooks, job scripts, and deploy scripts used to build the SakThai family — LoRA fine-tune → GGUF conversion → HF publish, all on free Kaggle GPUs. Part of the House of Sak.

---

Dataset Info

Attribute	Value
Files	16 (14 content files + README.md + .gitattributes)
Total size	~141 KB
Created	2026-07-06
Last updated	2026-07-30
Downloads	184
License	Apache 2.0
Repository type	Dataset (notebook & script collection)
Format	.py (PEP 723 inline-script), .ipynb (Jupyter notebook)

---

Data Fields

This dataset is a collection of Python scripts and Jupyter notebooks, not a tabular dataset. Each file is a self-contained artifact with the following attributes:

Field	Type	Description
filename	string	File name (e.g., job-0.5b-v7.py)
path	string	Full path within the repo (e.g., scripts/train-sakthai-1.5b-v2.py)
size	integer	File size in KB
type	string	Category: training-notebook, training-script, eval-script, validation-script, deployment-script, infra
model_target	string	Target model size: 0.5B, 1.5B, 7B, or all
purpose	string	One-sentence description of the file's role in the pipeline
era	string	Pipeline generation: v6-era, v7-era, or infra
dependencies	string	Key Python package dependencies (e.g., transformers, trl, peft, huggingface_hub)

Category Breakdown

Category	Count	Files
Training notebooks	2	sakthai-engine.ipynb, sakthai-7b-engine.ipynb
0.5B job scripts	6	train-sakthai-0.5b-v2.py, job-0.5b-v7.py, job-0.5b-hfjobs.py, job-0.5b-exp.py, job-0.5b-nanguard.py, job-0.5b-nanhunt.py
1.5B / 7B scripts	2	scripts/train-sakthai-1.5b-v2.py, scripts/sakthai-7b-post-train.py
Evaluation & validation	3	eval-bfcl-0.5b.py, validate.py, validate_exp.py
Deployment	1	deploy-endpoint.py
Infrastructure	2	README.md, .gitattributes

File Details

Training Notebooks

File	Size	Model	Era	Dependencies
sakthai-engine.ipynb	10.5 KB	1.5B (Qwen2.5-1.5B-Instruct)	v6	transformers, trl, peft, bitsandbytes, accelerate, huggingface_hub, wandb
sakthai-7b-engine.ipynb	24.0 KB	7B (Qwen2.5-7B-Instruct)	v6	transformers, trl, peft, bitsandbytes, accelerate, huggingface_hub, wandb

Job Scripts (0.5B)

File	Size	Model	Era	Dependencies
train-sakthai-0.5b-v2.py	6.8 KB	0.5B (Qwen2.5-0.5B-Instruct)	v6	transformers, trl, peft, datasets, huggingface_hub
job-0.5b-v7.py	10.2 KB	0.5B	v7	transformers, trl, peft, datasets, huggingface_hub
job-0.5b-hfjobs.py	7.8 KB	0.5B	v7	transformers, trl, peft, datasets
job-0.5b-exp.py	12.9 KB	0.5B	v7	transformers, trl, peft, datasets, wandb
job-0.5b-nanguard.py	13.2 KB	0.5B	v7	transformers, trl, peft, datasets
job-0.5b-nanhunt.py	11.8 KB	0.5B	v7	transformers, trl, peft, datasets

Training Scripts (1.5B, 7B)

File	Size	Model	Era	Dependencies
scripts/train-sakthai-1.5b-v2.py	4.5 KB	1.5B	v6	transformers, trl, peft, datasets
scripts/sakthai-7b-post-train.py	13.7 KB	7B	v6	transformers, peft, huggingface_hub, llama-cpp-python

Evaluation & Validation

File	Size	Model	Era	Dependencies
eval-bfcl-0.5b.py	3.9 KB	0.5B	v7	datasets, transformers, torch
validate.py	4.9 KB	0.5B	v7	datasets, transformers
validate_exp.py	3.7 KB	0.5B	v7	datasets, transformers

Deployment & Infra

File	Size	Purpose	Dependencies
deploy-endpoint.py	3.1 KB	Deploy to HF Inference Endpoints	huggingface_hub
.gitattributes	2.4 KB	Git LFS config	—
README.md	~7 KB	Dataset card (this file)	—

---

How These Notebooks Were Collected & Filtered

Origin. These notebooks were written from scratch by the SakThai project as part of the iterative model development pipeline. They are not scraped or collected from external sources — each script was authored to solve a specific training, evaluation, or deployment need.

Evolution. The collection grew organically across multiple training cycles:

1. v6 era (July 2026) — Initial notebooks trained the 1.5B and 7B models on sakthai-combined-v6 via Kaggle's free T4 GPU. These produced the first merged models and established the LoRA → GGUF pipeline.

2. v7 era (late July 2026) — A new generation of job scripts was written to fine-tune the 0.5B variant on sakthai-combined-v7. These scripts incorporated lessons from earlier failures:

   • No mid-run Hub pushes (save_strategy="no") — a checkpoint push failure had killed a 2-hour run
   • No Trackio logging (report_to="none") — Trackio's config-to-parquet export couldn't serialise PEFT's empty rank_pattern struct
   • Bench-exclusion filtering — every row reserved by sakthai-bench-v1 is excluded from training
   • Prompt masking — loss computed only on completion tokens, not the system prompt

3. Validation scripts — Two validate*.py scripts were added to prove every non-GPU component locally before launching paid jobs, catching data pipeline bugs early.

4. Experimental scripts — job-0.5b-exp.py, job-0.5b-nanguard.py, and job-0.5b-nanhunt.py capture specific run configurations for reproducibility of ablation studies on prompt masking and LoRA strategies.

Filtering criteria. Scripts were included if they:

• Directly contribute to training, evaluating, validating, or deploying a SakThai model
• Are self-contained (PEP 723 inline script metadata for dependencies where applicable)
• Are licensed Apache 2.0
• Are functionally distinct (duplicate-mode scripts like nanguard/nanhunt are included because they capture specific run configurations for reproducibility)

---

Usage

Load with the 🤗 Datasets library

from datasets import load_dataset

# Load the dataset (returns a Dataset with file metadata)
ds = load_dataset("Nanthasit/sakthai-kaggle-notebooks", split="train")
print(ds)
# Dataset({
#     features: ['filename', 'path', 'size', 'type', 'model_target', 'purpose', 'era', 'dependencies'],
#     num_rows: 16
# })

# List all filenames in the collection
for row in ds:
    print(f"{row['filename']:40s} {row['type']:20s} {row['model_target']:6s} {row['era']}")

List all available files via HfApi

from huggingface_hub import HfApi

api = HfApi()
files = api.list_repo_files("Nanthasit/sakthai-kaggle-notebooks", repo_type="dataset")
for f in sorted(files):
    print(f)

Download a specific script

from huggingface_hub import hf_hub_download

# Download a specific training script
script_path = hf_hub_download(
    repo_id="Nanthasit/sakthai-kaggle-notebooks",
    repo_type="dataset",
    filename="job-0.5b-v7.py"
)
print(f"Downloaded to: {script_path}")

Run a training job (PEP 723 inline-script)

# All job scripts are PEP 723 inline-script compatible
cd /tmp
curl -OL https://huggingface.co/datasets/Nanthasit/sakthai-kaggle-notebooks/raw/main/job-0.5b-v7.py
uv run job-0.5b-v7.py

Open notebooks in Kaggle

curl -OL https://huggingface.co/datasets/Nanthasit/sakthai-kaggle-notebooks/raw/main/sakthai-engine.ipynb
# Then upload to Kaggle via kaggle kernels push

---

File Inventory

#	File	Type	Size
1	sakthai-engine.ipynb	Training notebook (1.5B)	10.5 KB
2	sakthai-7b-engine.ipynb	Training notebook (7B)	24.0 KB
3	train-sakthai-0.5b-v2.py	Training script (0.5B, v6)	6.8 KB
4	job-0.5b-v7.py	Training script (0.5B, v7)	10.2 KB
5	job-0.5b-hfjobs.py	Training script (0.5B, HF Jobs)	7.8 KB
6	job-0.5b-exp.py	Experimental script (0.5B)	12.9 KB
7	job-0.5b-nanguard.py	Ablation: nan guard	13.2 KB
8	job-0.5b-nanhunt.py	Ablation: nan hunt	11.8 KB
9	scripts/train-sakthai-1.5b-v2.py	Training script (1.5B)	4.5 KB
10	scripts/sakthai-7b-post-train.py	Post-training (7B)	13.7 KB
11	eval-bfcl-0.5b.py	BFCL evaluation	3.9 KB
12	validate.py	Validation script	4.9 KB
13	validate_exp.py	Experimental validation	3.7 KB
14	deploy-endpoint.py	Deploy to HF Endpoints	3.1 KB
15	.gitattributes	Git LFS config	2.4 KB
16	README.md	This dataset card	~7 KB

---

Related Datasets & Models

Asset	Description
Nanthasit/sakthai-combined-v6	Base training data (v6 generation)
Nanthasit/sakthai-combined-v7	Extended training data (v7 generation)
Nanthasit/sakthai-bench-v1	Benchmark holdout data
Nanthasit/sakthai-tool-calling-v1	Tool-calling training data (SimpleToolCalling)
🤖 SakThai Agents	All models in the House of Sak
🏠 SakThai Model Family	Complete model family collection

---

_{Built with ❤️ by Beer · Part of the House of Sak — one family, one home.}