guava-v13b-qwen3.5-4b-red-basket
A Qwen3.5-VL 4B model for the red_objects_in_basket tabletop manipulation task,
from the Guava v13b model family. Full-weight inference export of checkpoint 50.
Related: guava-v13b-qwen3.5-4b · guava-v13b-qwen3.5-4b-no-counterfactual · guava-v13b-qwen3.5-4b-set-table
Status — read before using
Published as exported, with no evaluation run here.
- No evaluation results. No success rates, no held-out scores.
- The task designation is the uploader's. The export carried no run metadata
identifying its training data or split — the
red-basketlabel is not independently verifiable from these files. - Step 50, an early checkpoint, not selected against a validation metric.
- Untested load path. Requires
transformers5.8.1 (model_type: qwen3_5).
red_objects_in_basket is one of the 16 tasks in the Guava v13b dataset
(200 episodes: 134 main / 66 branch).
Model
| Architecture | Qwen3_5ForConditionalGeneration (vision + text) |
| Precision | bfloat16 |
| Hidden size | 2560, 32 layers (linear attention, full attention every 4th) |
| Vocab | 248,320 · Max positions |
| Vision | 24-layer ViT, patch 16, merge 2 |
| Shards | 2 safetensors, 8.47 GiB (224 + 499 tensors) |
Config, tokenizer, and chat template are byte-identical across this model family; only the weight shards differ.
Usage
from transformers import AutoModelForImageTextToText, AutoProcessor
model = AutoModelForImageTextToText.from_pretrained(
"AIcell/guava-v13b-qwen3.5-4b-red-basket", dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained("AIcell/guava-v13b-qwen3.5-4b-red-basket")
Coordinates are table-aligned (tabletop is z = 0) — do not apply table-height normalization a second time.