psycoplankton-14/Lumma-0.6B-SFT-4

Lumma-0.6B-SFT-4

SFT fine-tune of FrontiersMind/Lumma-0.6B-Base on a ToolACE + IndicAlign curriculum, trained with TRL's SFTTrainer (run lumma-sft-4).

These are the weights from step 1040 (best checkpoint by eval_loss = 1.0402).

2 epochs, lr 2e-5 cosine schedule. The final checkpoint (step 1040) is also the best by evalloss -- training had essentially converged by step 832 (evalloss 1.0399/1.0401/1.0402 at steps 832/936/1040).

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-4", 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