qwen3.5-9b-agentnet-cot-l2-step100

Training format (OpenCUA L2)

markdown
Copy code
## Thought:
<reasoning>

## Action:
<one-sentence>

## Code:
pyautogui.click(x=0.5, y=0.5)

Coordinates normalized to [0, 1]. The ## markdown headers help the base model emit the schema reliably (vs. the legacy bare Thought: form). See insagur/qwen3.5-9b-agentnet-ubuntu-1epoch for the legacy-format variant.

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
• Steps: 100 (preempted; 1 epoch = 300 steps)
• EMA teacher: target=block, decay=0.9995, α=0.5
• Sequence length: 3072
• Image tokens: 2048 (≈1.6M pixel cap)
• Save frequency: every 50 steps

Metrics @ step 100

Already approaches the fully-trained legacy-format model's eval loss (0.4622) at only 33% of training, suggesting the ## format converges faster.

Data

scripts/convert_agentnet_cot.py --cot_level l2 produces this format from AgentNet 5K trajectories with the same quality filter as the legacy converter (alignment≥7, efficiency≥5).

Inference

python
Copy code
from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "insagur/qwen3.5-9b-agentnet-cot-l2-step100",
    torch_dtype="bfloat16",
).to("cuda")
processor = AutoProcessor.from_pretrained("insagur/qwen3.5-9b-agentnet-cot-l2-step100")

system = (
    "You are a computer-use agent operating a Linux desktop. "
    "Respond using the OpenCUA L2 format:\n"
    "## Thought:\n<reasoning>\n\n## Action:\n<one-sentence>\n\n## Code:\n<pyautogui code with normalized [0,1] coords>"
)
# ... see scripts/eval.py in the training repo for full inference loop ...

Recipe

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

bash
Copy code
python scripts/convert_agentnet_cot.py --src ... --images_dir ... --out_dir ./agentnet_l2 --cot_level l2

CONFIG=configs/qwen35_9b_agentnet.yaml RUN_NAME=a100-9b-1ep-cot-l2 \
  sbatch --gpus=8 scripts/slurm_train_qwen.sbatch \
    data.train_jsonl=./agentnet_l2/train.jsonl \
    data.val_jsonl=./agentnet_l2/val.jsonl