OLMo 3 with SiameseNorm and DepthAttention
This repository is one of five standalone OLMo 3 checkpoints trained with SiameseNorm and DepthAttention. All model, configuration, tokenizer, and custom modeling files are stored directly at the repository root.
| Repository suffix | Training phase | Context length |
|---|---|---|
stage1 | Pretraining | 8,192 |
stage2 | Mid-training | 8,192 |
stage3 | Long-context training | 65,536 |
stage4-think | Think SFT | 65,536 |
stage4-instruct | Instruct SFT | 65,536 |
Stage 3 and Stage 4 apply YaRN only to Full-attention layers. Sliding-window attention layers retain the original RoPE and a 4,096-token window.
Loading
Set repo_id to the desired repository from the table above:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
repo_id = "ArchSpace-Collection/OLMo3-SiameseNorm-DepthAttention-<stage>"
tokenizer = AutoTokenizer.from_pretrained(
repo_id,
trust_remote_code=True,
fix_mistral_regex=False,
)
model = AutoModelForCausalLM.from_pretrained(
repo_id,
trust_remote_code=True,
dtype=torch.bfloat16,
attn_implementation="sdpa",
)
fix_mistral_regex=False preserves the tokenizer behavior used during
training. SDPA is the recommended and release-validated BF16 inference backend.
Architecture
- 16 transformer layers
- hidden size 2,048
- intermediate size 8,192
- 16 query heads and 16 key/value heads
- 128-dimensional attention heads
- 3:1 sliding-window/full-attention pattern
- 4,096-token sliding window
- reordered RMSNorm, SiameseNorm, and DepthAttention