Modelpublic

TrajectoryLabs/Nemotron-3.5-Lightning-Tau3-Legacy-LoRA

sourceHugging Faceotherupdated 24d agoView on Hugging Face

0likes37downloads

Model Card

Nemotron 3.5 Lightning — Tau3 Legacy LoRA

This repository contains the earlier Trajectory Tau3 LoRA checkpoint for NVIDIA Nemotron 3.5 Lightning 30B-A3B. It is published as the historical comparison point for our harness-matched retraining work.

Model details

—Base model: `nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16`, revision a9904d24bcc1d289a1950fa9d2b978c47cf903b9
—Format: PEFT LoRA adapter, rank 32, alpha 32, all linear layers
—Training checkpoint: step 80 from Trajectory experiment 1000218
—Intended use: research and evaluation of multi-turn tool-use agents, especially Tau3

The adapter does not include the base-model weights. Load the NVIDIA base model at the revision above, then apply this adapter with PEFT. Follow the base model's runtime and chat-template requirements.

python

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained(
    "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
    revision="a9904d24bcc1d289a1950fa9d2b978c47cf903b9",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(
    base,
    "TrajectoryLabs/Nemotron-3.5-Lightning-Tau3-Legacy-LoRA",
)

Limitations and license

Performance is sensitive to the agent harness, tools, user simulator, decoding settings, turn budget, and grading implementation. Compare models only under an identical protocol. This adapter is derived from NVIDIA Nemotron 3.5 Lightning and is distributed subject to the NVIDIA Open Model License Agreement. Review the base model card and license before use.