Mudit666/tiny-aya-global-qint4
Text Generation • 2B • Updated • 23
Tiny Aya 4-bit compression: models
qint4 Tiny Aya Global checkpoints from our safety-aware compression study. Code: https://github.com/Muditkumar123/tiny-aya-unlearning-4bit
Note plain-q4: every linear layer qint4.
Mudit666/tiny-aya-global-qint4-random77-bf16
Text Generation • 2B • Updated • 20Note random30-q4: 77 random linear layers kept in bf16 (size control).
Mudit666/tiny-aya-global-qint4-refusalmask77-bf16
Text Generation • 2B • Updated • 28Note protected-q4: 77 attribution-selected linear layers kept in bf16.
Mudit666/tiny-aya-tofu-c0v2-qint4
Text Generation • 2B • Updated • 21Note C1: the TOFU fine-tune C0v2 with every decoder linear qint4 (compression control for Study U).
Mudit666/tiny-aya-tofu-c0v2
Text Generation • 3B • Updated • 19Note C0v2: Tiny Aya Global fine-tuned on TOFU (bf16), the model unlearning starts from.
Mudit666/tiny-aya-tofu-npo-c2
Text Generation • 3B • Updated • 18Note C2: C0v2 after NPO unlearning of TOFU forget10 with SafeTune (bf16).
Mudit666/tiny-aya-tofu-npo-c2-qint4
Text Generation • 2B • Updated • 23Note C3: C2 with every decoder linear qint4 (the compressed unlearned model in Study U).