Fine-tunes Qwen3.5-35B for command execution.
# Set your HF token
export HF_TOKEN=hf_xxxxx
# Run training
python train.py --steps 50
# Or with custom parameters
python train.py --steps 100 --lr 5e-5 --batch 2
| Parameter | Default | Description |
|---|---|---|
| --steps | 50 | Training steps |
| --lr | 1e-4 | Learning rate |
| --batch | 1 | Batch size |
| --output | /tmp/nexus_model | Output directory |
| --local-only | false | Don't push to Hub |
| Variable | Required | Description |
|---|---|---|
| HF_TOKEN | Yes | HuggingFace token |
| OUTPUT_REPO | No | Output repo (default: Gerlai24/nexus-35b-execution) |
Base model: unsloth/Qwen3.5-35B-A3B-GGUF
Trained adapter will be pushed to HuggingFace Hub.
This space uses ZeroGPU for free GPU access.