OLMo3-3B-SiameseNorm-DepthAttention-stage3

OLMo 3 3B SiameseNorm + DepthAttention — Stage 3 Long-context Training

This repository is the Hugging Face export of o3sd3b-s3-s65536-g64-m1-ga2-tp2-cp8-dp32-h32-b2-lr2p5e4-w200-save1000-512npu-ptdata-0904062425-s3v1 at iteration 11921. This model preserves the trained SiameseNorm + DepthAttention architecture through bundled Hugging Face remote code. Load it with trust_remote_code=True.

• Training sequence length: 65,536
• Model context capacity: 65,536
• 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-stage3"
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.