team-nexus-gemma-4-e4b-it-text

The text-only language model of google/gemma-4-E4B-it (commit ee0ef6023621cff504d758262d4e04895a5af4a2), used by Team Nexus in the SinhalaMMLU shared task.

• The vision and audio towers (vision_tower, audio_tower, embed_vision, embed_audio) are removed.
• Every language-model tensor is copied unchanged, and the extraction script checks that each tensor equals its source.
• There is no fine-tuning.
• Class Gemma4ForCausalLM, with 7,463,013,376 parameters (bf16). Details are in SOURCE.json.
• It was built with training/scripts/make_text_only.py in SinhalaMMLU-SharedTask/team-nexus: python scripts/make_text_only.py google/gemma-4-E4B-it --revision ee0ef6023621cff504d758262d4e04895a5af4a2.

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("SinhalaMMLU-SharedTask/team-nexus-gemma-4-e4b-it-text")
model = AutoModelForCausalLM.from_pretrained("SinhalaMMLU-SharedTask/team-nexus-gemma-4-e4b-it-text", dtype="bfloat16")

License: Apache 2.0, the same as the original Gemma 4 model.