11-47/kaggle_llm_master_v2

Single split, 25,000 examples. Recommended random split for SFT:

• train: 23,500
• validation: 1,500

Stratify by category if you want balanced accelerator coverage.

Category breakdown: Total: 20,778 T4 dual / 4,222 TPU v3-8

Models covered: Llama-3.1-8B, Mistral-7B-v0.3, Qwen2.5-7B, Gemma-2-9b, Phi-3-medium, Hermes-3, DeepSeek-LLM-7B, Yi-1.5-9B, Flan-T5-XXL, Pythia-2.8B

Tasks: SFT, QLoRA 4bit, DPO, ORPO, continued pretraining, classification, embedding training, vision-text finetuning

---

Usage

Load with datasets

SFT training – QLoRA on Kaggle T4x2

Training a 7B base on this dataset with QLoRA r=64 fits comfortably in a single Kaggle T4 (16GB). For faster training, use both T4s with Accelerate: accelerate launch --multi_gpu --num_processes=2 train.py

Prompt format

Alpaca-style:

Proven formulas encoded in every example

T4 dual bring-up Unsloth QLoRA TPU v3-8 Kaggle API agentic Memory – 7B QLoRA on T4 ∼4.4GB base (NF4) + 0.2GB adapters + 2GB activations ≈ 6.7GB → fits 1x T4

Effective batch global_batch = per_device_batch × grad_accum_steps × world_size

---

Limitations

• Synthetic expert-curated. All 25,000 examples are programmatically generated from 100+ competition-proven templates, grounded in public Kaggle docs / kernels (Dec 2025 – May 2026). Not scraped from private notebooks. No PII, no secrets.
• Kaggle-specific. Paths, quotas, NCCL env vars, and API commands target Kaggle Notebooks (July 2026). Adapt for other platforms.
• Code is illustrative. Always review before running in production. Check transformers, peft, trl, unsloth, torch_xla versions – Kaggle images change.
• English only.

Intended use: SFT / instruction-tuning LLMs to generate correct Kaggle T4 / TPU training code, debug DDP/TPU jobs, and drive the Kaggle API CLI agentically.

---

Citation

---

License

Apache-2.0 – commercial use allowed.

---

Acknowledgments

Built from public sources:

• Kaggle Docs – TPUs / GPUs – https://www.kaggle.com/docs/tpu
• PyTorch Distributed Data Parallel – Kaggle T4x2 guide – LearnOpenCV
• Unsloth – Fine-Tuning Qwen VL on a Single T4 – Towards AI, May 2026
• Kaggle API – kaggle kernels push --accelerator NvidiaTeslaT4 – https://github.com/Kaggle/kaggle-api
• TRL – SFTTrainer with UnslothVisionDataCollator

Thanks to the Kaggle community for publishing competition kernels that made the ground-truth formulas possible.