kvn420/tenro-computer-use-v2-lora

TENRO-Ω++ Computer-Use v2 (7B VLM)

Agentic computer-use VLM for TENRO-Ω++, fine-tuned from Qwen/Qwen2.5-VL-7B-Instruct with QLoRA on the tenro-cuv2-dataset (4 206 samples) to emit structured action-JSON (click / type / scroll / launch / key / wait / finish …).

Training

• Method: QLoRA (4-bit NF4), LoRA r=16, alpha=32, dropout 0.05
• Target modules: q_proj, k_proj, v_proj, o_proj, up_proj, down_proj, gate_proj
• Steps: 150, max_seq_len 2048, grad accumulation 4
• Final train_loss: 1.381 (last loss 1.118), 0.143 epoch

Contents

path	description
tenro_computer_use_v2_lora/	LoRA adapter (final) + checkpoint-50/100/150

• fused/ — merged model weights (base + LoRA), 4-bit, model.safetensors
• tenro_base/ — TENRO base config/tokenizer (kvn420/tenro-omega-plusplus-7b-r2)

Usage (LoRA)

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
from peft import PeftModel
import torch

base = "Qwen/Qwen2.5-VL-7B-Instruct"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    base, torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(model, "kvn420/tenro-computer-use-v2-lora",
                                  subfolder="tenro_computer_use_v2_lora")
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")

Training lineage

• Kaggle kernel: dereckdeeds/tenro-cuv2-train (v4, COMPLETE)
• Package: tenro/transcendant/computer_use_v2