qwen3.5-9b-agentnet-ubuntu-1epoch

Training format

Legacy Thought: / Action: / Code: template (no ## headers). For the OpenCUA ## format variant see insagur/qwen3.5-9b-agentnet-cot-l2-step100.

markdown
Copy code
Thought: <reasoning>
Action: <one-sentence>
Code:
pyautogui.click(x=0.5, y=0.5)

Coordinates are normalized to [0, 1] of screen width/height.

Training config

• Hardware: 1 × 8 A100 80GB SXM4
• Distributed: DeepSpeed ZeRO-2 + bf16
• Optimizer: AdamW, LR 1e-5 cosine, warmup 200 steps
• Batch: per_device_bs=1 × grad_accum=16 × 8 GPU = global batch 128
• Epochs: 1 (300 steps)
• EMA teacher: target=block (last ViT block), decay=0.9995, α=0.5
• Sequence length: 3072 (truncated; p99=2713)
• Image tokens: 2048 (≈1.6M pixel cap; ~1689×950 post-resize)
• Gradient checkpointing: on
• Train runtime: 5h 14m

Metrics (final)

Offline eval on 50 val samples (scripts/eval.py):

Format adherence is the main limitation — the legacy bare-prefix format isn't reliably emitted by Qwen3.5; the OpenCUA ## variant addresses this.

Data

AgentNet Ubuntu 5K trajectories filtered with task_completed AND alignment≥7 AND efficiency≥5 and per-step last_step_correct AND NOT last_step_redundant. 5% trajectory-level val split.

Inference

python
Copy code
from transformers import AutoModelForImageTextToText, AutoProcessor
from PIL import Image

model = AutoModelForImageTextToText.from_pretrained(
    "insagur/qwen3.5-9b-agentnet-ubuntu-1epoch",
    torch_dtype="bfloat16",
).to("cuda")
processor = AutoProcessor.from_pretrained("insagur/qwen3.5-9b-agentnet-ubuntu-1epoch")

system = (
    "You are a computer-use agent operating a Linux desktop. "
    "You receive the user's task and the current screenshot. "
    "Respond with your reasoning, the action description, and the pyautogui code to execute. "
    "All coordinates are normalized to [0, 1] of screen width/height. "
    "Format your response exactly as:\n"
    "Thought: <your reasoning>\nAction: <one-sentence description of the next action>\nCode:\n<pyautogui code>"
)
image = Image.open("screenshot.png").convert("RGB")
messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": "Task: Open the terminal.\n<image>\nWhat is the next action?"},
    ]},
]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=256)
print(processor.batch_decode(out, skip_special_tokens=True)[0])

Recipe

Training code: https://github.com/2bhapby/gui_internal_worldmodel

bash
Copy code
SMOKE=0 WANDB=1 DS=z2 NPROC=8 PER_DEVICE_BS=1 GRAD_ACCUM=16 \
RUN_NAME=a100-full-9b-1epoch \
sbatch --gpus=8 scripts/slurm_train_qwen35_4b_agentnet.sbatch \
  optim.num_train_epochs=1