---
library_name: transformers
pipeline_tag: fill-mask
tags: [gpt-bert, babylm, remote-code]
license: other
---
# jumelet/gptbert-pol-250steps-base

GPT-BERT style BabyBabyLLM model for language **pol**.

This repository may include both *main* and *EMA* variants.

**Default variant exposed to generic loaders:** `ema`

## Variants Available
ema, main

## Files
- model.safetensors (alias of default variant)
- model_ema.safetensors
- pytorch_model.bin (legacy PyTorch format)
- pol-2gpu-250steps.bin (raw training checkpoint)
- pol-2gpu-250steps_ema.bin (raw training checkpoint)

## Configuration
```json
{
  "attention_probs_dropout_prob": 0.1,
  "hidden_dropout_prob": 0.1,
  "hidden_size": 768,
  "intermediate_size": 2560,
  "max_position_embeddings": 512,
  "position_bucket_size": 32,
  "num_attention_heads": 12,
  "num_hidden_layers": 12,
  "vocab_size": 16384,
  "layer_norm_eps": 1e-05,
  "force_causal_mask": true,
  "classifier_dropout": 0.1,
  "classifier_layer_norm_eps": 1e-05,
  "num_labels": 2
}
```
Tokenizer file: `tokenizer_pol_vs16384.json`

## Quick Usage
```python
from transformers import AutoTokenizer, AutoModelForMaskedLM
model_id = 'jumelet/gptbert-pol-250steps-base'
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForMaskedLM.from_pretrained(model_id, trust_remote_code=True)
out = model(**tok('Hello world', return_tensors='pt'))
```

### Forced Causal Attention
Causal attention is enforced during inference by applying a triangular future mask inside the remote code.
This prevents the hybrid GPT-BERT layers from attending to future tokens even when a bidirectional mask is provided.

### Sequence Classification
`GPTBertForSequenceClassification` mirrors the original GLUE classifier head for downstream fine-tuning.
```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
model_id = 'jumelet/gptbert-pol-250steps-base'
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id, trust_remote_code=True)
outputs = model(**tok('This movie was great!', return_tensors='pt'))
print(outputs.logits)
```

## Notes
- Converted on 2025-10-07T01:14:48.240581+00:00
- Weights are the exact trained parameters; no new layers were initialized.
- Requires `trust_remote_code=True` due to custom architecture.