You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Llama-3.2-1B — OASIS (softmax1)

OASIS softmax1 变体（eager attention + history propagation）（attn_softmax=softmax1, attn_res_softmax_fn=softmax1）.

Base model: meta-llama/Llama-3.2-1B
Training data: BookCorpus + Wiki40B (en)
Steps: 1000 (warmup 100), block_size 512, effective batch 192 (6 × 4 GPU × 8 grad-accum)
Optimizer: AdamW, lr 4e-4, linear, weight_decay 0.1
Precision: fp16 mixed (fp32 master weights)

Research checkpoint from an outlier-efficiency study; only 1000 steps, not production-ready.

Companion run: robinzixuan/llama-3.2-1b-vanilla.

Safetensors

Model size

1B params

Tensor type

F32

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Base model

Finetuned

(938)

this model