psycoplankton-14/Lumma-0.6B-SFT-5-checkpoint-520

Lumma-0.6B-SFT-5-checkpoint-520

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 520 (best checkpoint by eval_loss = 0.9883).

3 epochs planned, lr 8e-5 constant schedule. Best checkpoint by evalloss was step 520 (end of epoch 1) -- evalloss got worse in epochs 2-3 (overfitting), so these weights are from epoch 1, not the final step.

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-checkpoint-520", 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