# haznitrama/babybabellm-gpt_bert-ace-causal

GPT-BERT style BabyBabyLLM monolingual model for language **ace**.

This repository mirrors the layout of the multi-all reference models: it may contain 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)

## Configuration
```json
{
  "attention_probs_dropout_prob": 0.1,
  "hidden_dropout_prob": 0.1,
  "hidden_size": 384,
  "intermediate_size": 1280,
  "max_position_embeddings": 512,
  "position_bucket_size": 32,
  "num_attention_heads": 6,
  "num_hidden_layers": 12,
  "vocab_size": 8192,
  "layer_norm_eps": 1e-05,
  "auto_map": {
    "AutoConfig": "configuration_gpt_bert.GPTBertConfig",
    "AutoModel": "modeling_gpt_bert.GPTBertForMaskedLM",
    "AutoModelForCausalLM": "modeling_gpt_bert.GPTBertForMaskedLM",
    "AutoModelForMaskedLM": "modeling_gpt_bert.GPTBertForMaskedLM"
  },
  "return_dict": true,
  "output_hidden_states": false,
  "torchscript": false,
  "dtype": "float32",
  "pruned_heads": {},
  "tie_word_embeddings": true,
  "chunk_size_feed_forward": 0,
  "is_encoder_decoder": false,
  "is_decoder": false,
  "cross_attention_hidden_size": null,
  "add_cross_attention": false,
  "tie_encoder_decoder": false,
  "architectures": [
    "GPTBertForMaskedLM"
  ],
  "finetuning_task": null,
  "id2label": {
    "0": "LABEL_0",
    "1": "LABEL_1"
  },
  "label2id": {
    "LABEL_0": 0,
    "LABEL_1": 1
  },
  "task_specific_params": null,
  "problem_type": null,
  "tokenizer_class": null,
  "prefix": null,
  "bos_token_id": null,
  "pad_token_id": null,
  "eos_token_id": null,
  "sep_token_id": null,
  "decoder_start_token_id": null,
  "max_length": 20,
  "min_length": 0,
  "do_sample": false,
  "early_stopping": false,
  "num_beams": 1,
  "num_beam_groups": 1,
  "diversity_penalty": 0.0,
  "temperature": 1.0,
  "top_k": 50,
  "top_p": 1.0,
  "typical_p": 1.0,
  "repetition_penalty": 1.0,
  "length_penalty": 1.0,
  "no_repeat_ngram_size": 0,
  "encoder_no_repeat_ngram_size": 0,
  "bad_words_ids": null,
  "num_return_sequences": 1,
  "output_scores": false,
  "return_dict_in_generate": false,
  "forced_bos_token_id": null,
  "forced_eos_token_id": null,
  "remove_invalid_values": false,
  "exponential_decay_length_penalty": null,
  "suppress_tokens": null,
  "begin_suppress_tokens": null,
  "_name_or_path": "",
  "transformers_version": "4.56.1",
  "tf_legacy_loss": false,
  "use_bfloat16": false,
  "model_type": "gpt_bert",
  "output_attentions": false
}
```
Tokenizer file: `tokenizer_ace_vs8192.json`

## Quick Usage
```python
from transformers import AutoTokenizer, AutoModelForMaskedLM
model_id = 'haznitrama/babybabellm-gpt_bert-ace-causal'
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForMaskedLM.from_pretrained(model_id, trust_remote_code=True)
out = model(**tok('Hello world', return_tensors='pt'))
```
Select a specific variant explicitly (when both present):
```python
# Load EMA weights explicitly if both are present
from safetensors.torch import load_file
import torch
from transformers import AutoConfig, AutoModelForMaskedLM
model_id = 'haznitrama/babybabellm-gpt_bert-ace-causal'
config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForMaskedLM.from_config(config, trust_remote_code=True)
state_dict = torch.load('pytorch_model.bin')  # or load_file('model_ema.safetensors')
model.load_state_dict(state_dict, strict=False)
```

### Causal LM Wrapper
This repo includes a lightweight GPTBertForCausalLM wrapper.
Generation example:
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
mid='haznitrama/babybabellm-gpt_bert-ace-causal'
tok=AutoTokenizer.from_pretrained(mid)
model=AutoModelForCausalLM.from_pretrained(mid, trust_remote_code=True)
print(tok.decode(model.generate(**tok('Hello', return_tensors='pt'), max_new_tokens=20)[0], skip_special_tokens=True))
```

## Notes
- Converted on 2025-09-16T06:15:08.548402Z
- Safe serialization (safetensors) used; `pytorch_model.bin` added for legacy tools.
- Requires `trust_remote_code=True` due to custom architecture.
- EMA (Exponential Moving Average) weights can yield slightly better evaluation metrics; choose according to your needs.