qwen3-vl-32b-soccer-v11-lora

LoRA adapter (checkpoint-150, the production pick) for the soccer event classifier. ~2.27 GB. Use this if you want to:

Apply the adapter on top of a different base (e.g., the bf16 Qwen3-VL-32B rather than the FP8 quant)
Continue training from this checkpoint
Inspect the LoRA weights

For inference, most users will want the merged FP8 model instead: acatorcini/qwen3-vl-32b-soccer-v11-fp8. That single artifact loads directly into vLLM with no merge step.

Adapter config

peft_type: LORA
r: 64
lora_alpha: 128
lora_dropout: 0.05
target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
base_model_name_or_path: (an internal v8-32b BF16 merge — see below)

Base-model note: The adapter's adapter_config.json references an internal predecessor checkpoint (v8-32b/merged-bf16-c300) as its base. The merged FP8 production model was built by applying THIS adapter to THAT predecessor base and then quantizing. To reproduce the merge against the public Qwen3-VL-32B base, expect slight numerical differences.

Use with PEFT

from transformers import AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-VL-32B-Instruct",
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "acatorcini/qwen3-vl-32b-soccer-v11-lora")