Finetuned from Unsloth/Qwen3-4B-Instruct-2507 using QLoRA + Unsloth. Trained to respond directly without chain-of-thought (<think> blocks stripped from all training data).

What's New in V2.1: The best I can afford on free Kaggle GPU

V2.1 further expands the training data to 7 datasets covering Gemini, Claude, and GPT-sourced reasoning traces, and was trained on Kaggle's dual T4 GPU environment (though only one GPU was used for training). Dataset size after deduplication contains ~4,752 examples with MinHash-based similarity filtering at 0.4 Jaccard threshold.

Training Data

Dataset	Source Model
TeichAI/gemini-3-pro-preview-high-reasoning-250x	Gemini
TeichAI/gemini-3-pro-preview-high-reasoning-1000x	Gemini
Roman1111111/gemini-3.1-pro-hard-high-reasoning	Gemini
crownelius/Opus-4.6-Reasoning-3300x	Claude
TeichAI/claude-4.5-opus-high-reasoning-250x	Claude
TeichAI/claude-sonnet-4.5-high-reasoning-250x	Claude
TeichAI/gpt-5.2-high-reasoning-250x	GPT

~7,322 rows loaded, ~4,752 after MinHash deduplication (threshold 0.4). <think> blocks stripped from all assistant turns before training.

Training Details

Parameter	Value
Method	QLoRA (4-bit NF4) + Unsloth
LoRA rank	16
LoRA alpha	16
Epochs	1
Steps	594
Batch size	2 per device × 4 gradient accumulation = 8 effective
Learning rate	1e-4 (cosine schedule)
Max seq length	2048
Optimizer	AdamW 8-bit
Hardware	Kaggle Tesla T4 (14.56 GB VRAM)
Training time	~7h 23min
Trainable params	33M / 4.05B (0.81%)
Peak VRAM	~10.2 GB

Training used train_on_responses_only — loss computed on assistant completions only.

Files

*.gguf — IQ4_XS quantized, ready for LM Studio / Ollama / llama.cpp
lora-adapter/ — Raw LoRA weights for merging with the base model

Usage (Ollama)

ollama run VladHong/Qwen3-4B-Instruct-NoThink-V2.1

License Note

Base model is Apache 2.0. Training data includes AI-generated content — review upstream dataset terms before commercial use.