Lumma-0.6B-SFT-5

Lumma-0.6B-SFT-5

SFT fine-tune of FrontiersMind/Lumma-0.6B-Base on a ToolACE + IndicAlign curriculum, trained with TRL's SFTTrainer (run lumma-sft-5).

These are the weights from step 1560 (best checkpoint by eval_loss = 1.0924).

3 epochs, lr 8e-5 constant schedule. These are the final weights (step 1560, end of epoch 3). Note this is not the best checkpoint by eval_loss -- eval_loss got worse in epochs 2-3 (overfitting) after bottoming out at 0.9883 (step 520, end of epoch 1); see psycoplankton-14/Lumma-0.6B-SFT-5-checkpoint-520 for that checkpoint.

Quick start

from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="psycoplankton-14/Lumma-0.6B-SFT-5", trust_remote_code=True, device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

This model uses a custom architecture ("nandi") and requires trust_remote_code=True.

Framework versions

• TRL: 1.8.0
• Transformers: 5.13.1
• Pytorch: 2.6.0+cu124
• Datasets: 4.8.3
• Tokenizers: 0.22.2