---
language:
  - en
license: apache-2.0
library_name: transformers
tags:
  - open-reason
  - causal-lm
  - cpu
datasets:
  - theworker02/open-reason
base_model: gpt2-scratch
---

# Open Reason XL (CPU)

GPT-2-style causal LM trained **from scratch** on the Open Reason SFT split.
Exact parameter count: **443,719,680**. This is **not** a 1B model and is **not**
`theworker02/open-reason-1b`.

## Training facts

- Parameters: 443,719,680 (`sum(p.numel() for p in model.parameters())`)
- Architecture: n_layer=22, n_embd=1280, n_head=20, vocab=8192, seq=256
- Steps: 120 (batch 1, gradient accumulation 2, gradient checkpointing)
- Device: **host CPU** (AMD Ryzen 9 9950X, 32 threads)
- `torch`: 2.12.0+cpu; `cuda_available=False`; Docker not installed
- NVIDIA CUDA was not used. AMD GPU / ROCm / DirectML were not used
- Dataset: [`theworker02/open-reason`](https://huggingface.co/datasets/theworker02/open-reason) pipeline **1.4.0**
- SFT rows: **3175** (`data/release/all.jsonl`)
- Final training NLL: **5.8116**
- License: Apache-2.0
- No held-out exact-match / coding / math benchmark scores are claimed

Related checkpoints (none of these is a 1B model):

- Dataset: https://huggingface.co/datasets/theworker02/open-reason
- Small: https://huggingface.co/theworker02/open-reason-small
- Medium: https://huggingface.co/theworker02/open-reason-medium
- Large: https://huggingface.co/theworker02/open-reason-large
- XL: https://huggingface.co/theworker02/open-reason-xl

No Reddit sources. Project license Apache-2.0.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("theworker02/open-reason-xl")
model = AutoModelForCausalLM.from_pretrained("theworker02/open-reason-xl")
```