AIcell/Guava-4B

guava-v13b-qwen3.5-4b

A Qwen3.5-VL 4B model fine-tuned on the Guava v13b robot-manipulation dataset (2,268 episodes across 16 tabletop tasks). This is the full-weight inference export of checkpoint 213 from the v13b-main-full-2707 run.

Training data: AIcell/guava-v13b (private).

Status — read before using

This checkpoint is published as exported. It has not been benchmarked here, and the following carry over from the training data and the export itself:

• No evaluation results. No success rates, no held-out scores. Nothing in this repo establishes how well the model performs on any task.
• One mid-training checkpoint. Step 213 of the run — not selected against a validation metric, because the dataset ships no held-out split.
• Inherited data caveat. The training data carries an unresolved physical-clearance issue on 05-24__push_cereal__trial_0308.
• Untested load path. The config requires transformers 5.8.1 (model_type: qwen3_5). Older versions will not recognize this architecture.

Model

Architecture	Qwen3_5ForConditionalGeneration (vision + text)
Precision	bfloat16
Hidden size	2560, 32 layers (linear attention, full attention every 4th)
Vocab	248,320
Max positions	262,144
Vision	24-layer ViT, patch 16, merge 2
Shards	2 safetensors, 8.47 GiB total

Usage

from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "AIcell/guava-v13b-qwen3.5-4b", dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained("AIcell/guava-v13b-qwen3.5-4b")

Episodes in the training data contain up to 30 images, and coordinates are already table-aligned (tabletop is z = 0) — do not apply table-height normalization a second time.

Tasks

apple_juice_order, bread_near_lemon, can_in_bin, close_drawer, cube_stack, cube_under_cup, hotdog_near_donut, milk_near_cup, open_drawer, pick_up_orange, push_basket, push_cereal, red_objects_in_basket, remove_cube_from_tray, shell_game, tomato_in_bowl