LLM-OS-Models/gemma-4-26B-A4B-it-Terminal-SFT-Native-Liquid-1Epoch

Summary

• Base model: google/gemma-4-26B-A4B-it
• Source dataset/cache: /home/work/.data/gemma4_native_sft/datasets/google__gemma-4-26B-A4B-it__liquid_raw_json_masked_8192
• Training format: Gemma 4 native chat template
• Labels: assistant JSON command response only
• Prompt/history labels are masked with -100
• Previous assistant thinking blocks are stripped from history

TB2-lite

• Result: pending

Notes

• Source checkpoint: /home/work/.data/gemma4_native_sft/models/google__gemma-4-26B-A4B-it__terminal_sft_native_liquid_2epoch/checkpoint-1020
• Checkpoint step: 1020
• Trainer epoch: 1.0000
• TB2-lite score: pending GPU evaluation
• Upload policy: checkpoint uploaded immediately after save; score card updates after evaluation.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("LLM-OS-Models/gemma-4-26B-A4B-it-Terminal-SFT-Native-Liquid-1Epoch")
model = AutoModelForCausalLM.from_pretrained("LLM-OS-Models/gemma-4-26B-A4B-it-Terminal-SFT-Native-Liquid-1Epoch", torch_dtype="auto")