OLMo3-3B-SiameseNorm-DepthAttention-stage2

OLMo 3 3B SiameseNorm + DepthAttention — Stage 2 Mid-training

This repository is the Hugging Face export of o3sd3b-s2-s8192-g256-m1-ga2-tp2-cp1-dp128-h16-b8-mc2-lr2p071e4-256npu-arch-0903100156-s2v2 at iteration 47684. This model preserves the trained SiameseNorm + DepthAttention architecture through bundled Hugging Face remote code. Load it with trust_remote_code=True.

• Training sequence length: 8,192
• Model context capacity: 8,192
• Sliding-window size: 4,096
• Attention pattern: [SWA, SWA, SWA, Full]
• Vocabulary: 100,278 real tokens; 74 Megatron padding-only rows removed

Stage 3/4 use the frozen 65,536-token configuration. YaRN applies to the Full Attention layers; SWA layers retain their original RoPE and 4,096-token local window.

Loading

Use transformers>=4.57.6,<5.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "ArchSpace-Collection/OLMo3-3B-SiameseNorm-DepthAttention-stage2"
tokenizer = AutoTokenizer.from_pretrained(
    repo_id,
    trust_remote_code=True,
    use_fast=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 exact tokenizer behavior used during training. Conversion provenance, per-tensor hashes, and CPU validation results are included in conversion_manifest.json, SHA256SUMS, and hf_validation_report.json.