---
library_name: transformers
pipeline_tag: text-generation
tags:
- olmo3
- baseline
- safetensors
- sliding-window-attention
---

# OLMo 3 1B Baseline — Stage 2 Mid-training

This repository contains the pure OLMo 3 1B baseline checkpoint from
`o3b1b-s2-s8192-g256-m1-tp1-cp2-dp256-hsdp32-b8-halo-norecomp-lr2p071235e4-save10000-512npu-share-20260801-v1` at iteration `47684`. SiameseNorm and Depth-Attention
are disabled.

- 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 rows removed

Stage 3/4 apply YaRN only to Full-Attention layers. OLMo 3 SWA layers use the
original RoPE and retain their 4,096-token local window.

## Loading

`transformers>=4.57.6,<5` is required.

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "ArchSpace-Collection/OLMo3-1B-SiameseNorm-DepthAttention-baseline-stage2"
tokenizer = AutoTokenizer.from_pretrained(
    repo_id,
    use_fast=True,
    fix_mistral_regex=False,
)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    dtype=torch.bfloat16,
    attn_implementation="sdpa",
)
```

`fix_mistral_regex=False` preserves the tokenizer behavior used for training.
The checkpoint uses the official Transformers `Olmo3ForCausalLM`
implementation and does not require remote code.