BEE-spoke-data/neobert-100k-test

NeoBERT Model

This is a NeoBERT model trained with pszemraj/NeoBERT and exported to transformers format.

Model Details

• Architecture: NeoBERT
• Hidden Size: 768
• Layers: 12
• Attention Heads: 12
• Vocab Size: 31999
• Max Length: 4096
• Dtype: float32

Usage

For Masked Language Modeling (Fill-Mask)

from transformers import AutoModelForMaskedLM, AutoTokenizer
import torch

repo_id = "BEE-spoke-data/neobert-100k-test"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained(repo_id, trust_remote_code=True)

# Example: Fill in masked tokens
text = "NeoBERT is the most [MASK] model of its kind!"

# Tokenize (handling Metaspace tokenizer's space tokens)
inputs = tokenizer(text, return_tensors="pt")
input_ids = inputs["input_ids"][0].tolist()

# Remove extra space tokens before [MASK] if present (Metaspace tokenizer quirk)
cleaned_ids = []
for i, token_id in enumerate(input_ids):
    if token_id == 454 and i < len(input_ids) - 1 and input_ids[i + 1] == tokenizer.mask_token_id:
        continue
    cleaned_ids.append(token_id)

if len(cleaned_ids) != len(input_ids):
    inputs["input_ids"] = torch.tensor([cleaned_ids])
    inputs["attention_mask"] = torch.ones_like(inputs["input_ids"])

# Get predictions
with torch.no_grad():
    outputs = model(**inputs)
    mask_pos = (inputs["input_ids"] == tokenizer.mask_token_id).nonzero(as_tuple=True)[1][0]
    predictions = outputs.logits[0, mask_pos].topk(5)

# Display top predictions
for idx, score in zip(predictions.indices, predictions.values):
    token = tokenizer.decode([idx])
    print(f"{token}: {score:.2f}")

For Embeddings / Feature Extraction

from transformers import AutoModel, AutoTokenizer

repo_id = "BEE-spoke-data/neobert-100k-test"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModel.from_pretrained(repo_id, trust_remote_code=True)

# Example: Generate embeddings
text = "NeoBERT is an efficient transformer model!"
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)

# Get CLS token embedding
cls_embedding = outputs.last_hidden_state[:, 0, :]
print(f"Embedding shape: {cls_embedding.shape}")

Training Configuration

Full Config (click to expand)

Full training config:

model:
  hidden_size: 768
  num_hidden_layers: 12
  num_attention_heads: 12
  intermediate_size: 3072
  max_position_embeddings: 4096
  vocab_size: 31999
  rope: true
  rms_norm: true
  hidden_act: swiglu
  dropout_prob: 0.05
  norm_eps: 1.0e-05
  embedding_init_range: 0.02
  decoder_init_range: 0.02
  classifier_init_range: 0.02
  flash_attention: true
  ngpt: false
  base_scale: 0.03227486121839514
  pad_token_id: 0
dataset:
  name: EleutherAI/SmolLM2-1.7B-stage-4-100B
  path: ''
  num_workers: 4
  streaming: true
  cache_dir: null
  max_seq_length: 1024
  validation_split: null
  train_split: train
  eval_split: train[:1%]
  num_proc: 8
  shuffle_buffer_size: 10000
  pre_tokenize: false
  pre_tokenize_output: null
  load_all_from_disk: false
  force_redownload: false
  pretraining_prob: 0.3
  min_length: 512
tokenizer:
  name: BEE-spoke-data/wordpiece-tokenizer-32k-en_code-msp
  path: null
  max_length: 1024
  padding: max_length
  truncation: true
  vocab_size: 31999
optimizer:
  name: adamw
  lr: 0.0001
  weight_decay: 0.01
  betas:
  - 0.9
  - 0.98
  eps: 1.0e-08
scheduler:
  name: cosine
  warmup_steps: 5000
  total_steps: null
  num_cycles: 0.5
  decay_steps: 50000
  warmup_percent: null
  decay_percent: null
trainer:
  per_device_train_batch_size: 16
  per_device_eval_batch_size: 16
  gradient_accumulation_steps: 4
  max_steps: 100000
  save_steps: 10000
  eval_steps: 5000
  logging_steps: 25
  output_dir: ./outputs/neobert_100m_100k
  overwrite_output_dir: true
  bf16: true
  gradient_checkpointing: false
  gradient_clipping: null
  mixed_precision: 'no'
  seed: 42
  resume_from_checkpoint: false
  disable_tqdm: false
  dataloader_num_workers: 0
  use_cpu: false
  report_to:
  - wandb
  tf32: true
  max_ckpt: 3
  train_batch_size: 16
  eval_batch_size: 32
datacollator:
  mlm_probability: 0.2
  pad_to_multiple_of: 8
wandb:
  project: neobert-pretraining
  entity: null
  name: neobert-100m-100k
  tags: []
  mode: online
  log_interval: 100
  resume: never
  dir: logs/wandb
task: pretraining
accelerate_config_file: null
mixed_precision: bf16
mteb_task_type: all
mteb_batch_size: 32
mteb_pooling: mean
mteb_overwrite_results: false
pretrained_checkpoint: latest
use_deepspeed: true
seed: 69
debug: false