saad1926q/Qwen3-4B-brainrot

🧠 unslothtrainingcheckpoints (Qwen3-4B fine-tuned on Brainrot)

This is a fine-tuned version of unsloth/Qwen3-4B-unsloth-bnb-4bit, trained using TRL (Transformers Reinforcement Learning) and Unsloth AI. The model was fine-tuned on the [Brainrot-xK-large dataset](https://huggingface.co/datasets/GoofyLM/Brainrot-xK-large) — a chaotic, Gen Z internet-flavored dataset full of memes, slang, sass, and semi-ironic wisdom.

It responds in a style that blends internet humor, casual language, and a touch of TikTok energy. Great for applications that need a little more vibe than traditional models.

🗣️ Example Outputs

🔹 Prompt: “summarize the story of one piece”

Response: <think>

</think>

omg, u literally need to read the whole thing, but here's a summary: the story is about a pirate named luffy, who wants to become the world's greatest pirate. he is a big fan of the "grand line" and is determined to find the legendary treasure. along the way, he meets a bunch of cool crew members and they have to go through a buncha challenges, but the main goal is to get the big treasure and become the "king of the pirates"!<|im_end|>

🔹 Prompt: “best programming language”

Response: <think>

</think>

lol ik u r a noob, but r u tryin 2 get better at coding?<|im_end|>

🧠 Training Details

The model was fine-tuned in two phases using SFT with Unsloth's optimized training approach:

Phase 1

• —Epochs: 45
• —Batch size: 2 (with gradient accumulation)
• —Learning rate: 2e-4
• —Optimizer: adamw_8bit
• —Scheduler: Linear
• —Seed: 3407
• —Precision: Full (fp16/bf16 disabled)

Phase 2

• —Epochs: 6 (continuation with same settings)

No evaluation dataset was used during fine-tuning. The dataset field used was "text".

Citations

Cite TRL as:

@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}