---
license: apache-2.0
language:
- en
library_name: transformers
pipeline_tag: text-generation
datasets:
- openbmb/Ultra-FineWeb
tags:
- moe
- booper
- pretraining
---

# Booper-Big

Booper-Big is a from-scratch English MoE language model trained on the English split of
`openbmb/Ultra-FineWeb`. It uses the standard Hugging Face Mixtral implementation.

| Property | Value |
|---|---:|
| Total parameters | 149,602,432 |
| Active parameters per token | 50,512,000 |
| Layers / width / heads | 6 / 896 / 14 |
| Experts / active experts | 7 / 1 |
| Context window | 4,096 |
| Tokenizer | Booper byte-BPE, 16,384 tokens |
| Pretraining tokens | 5,000,036,352 |
| Training precision | BF16 compute, FP32 master weights; BF16 published weights |
| Final training loss | 1.8922 |
| Held-out Ultra-FineWeb loss | 1.5720 |

The training stream filtered the already-curated English split to quality score ≥0.70 and held
the last prepared shard out of training. The model was trained with 1,024-token packed sequences;
RoPE and the published configuration support a 4,096-token context window.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ProCreations/Booper-Big"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto")
```

This is a small base model, not an instruction-following assistant. Outputs may be inaccurate or unsafe.