hawk_parfind_en_nl_zh_equal

hawk_parfind_en_nl_zh_equal

Multilingual Hawk causal language model for the BabyLM 2026 multilingual setting, paired with its own custom ParadigmFinder tokenizer trained on English, Dutch, and Chinese.

This package is staged for direct Hugging Face upload. Because both the model and tokenizer rely on custom runtime code, loading requires trust_remote_code=True.

Model

• architecture: HawkForCausalLM
• model type: hawk_rglru
• max positions: 1024
• exported model vocab size: 44091
• default main bundle source: chck_400M

Tokenizer

• class: EnhancedParadigmTokenizerWrapper
• paradigm family: ParadigmFinder
• multilingual budgeting: soft
• budget per language: 16384
• languages: eng, nld, zho
• space-free lexicon language: zho

This tokenizer includes the custom runtime files needed for HF loading:

• tokenizer.py
• boundary_discovery.py
• preprocessing.py
• paradigm_utils.py
• paradigms.json
• multilingual_meta.json
• modeling_hawk.py

Revisions

The following training checkpoints are prepared as self-contained HF revisions:

• chck_1M
• chck_2M
• chck_3M
• chck_4M
• chck_5M
• chck_6M
• chck_7M
• chck_8M
• chck_9M
• chck_10M
• chck_20M
• chck_30M
• chck_40M
• chck_50M
• chck_60M
• chck_70M
• chck_80M
• chck_90M
• chck_100M
• chck_200M
• chck_300M
• chck_400M

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

repo_id = "NeTSlab/hawk_parfind_en_nl_zh_equal"

tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo_id, trust_remote_code=True)

prompt = "The child looked at"
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=30)
print(tokenizer.decode(output[0]))

If your Hugging Face cache is not writable, set:

export HF_HOME=/tmp/hf_cache

Notes

• This is a BabyLM-scale research model, not a production general-purpose LM.
• The tokenizer and model both use custom remote code.
• The package preserves each checkpoint's own tokenizer.py and modeling_hawk.py so revision uploads stay aligned with the trained artifacts.
• main is intentionally pinned to chck_400M for this package refresh.