Hoshipu

stageguard-qwen3.5-0.8b-b1k-task69-removed-fullparam-45k

README

License: apache-2.0

Training

Final optimizer step: 45,000
Training schedule: initial 35,000 steps followed by 10,000 continuation steps
Training method: full-parameter supervised fine-tuning in bfloat16
Micro batch size: 4
Gradient accumulation: 1
Learning rate during the 35k-to-45k continuation: 1e-5
Gradient checkpointing: disabled
Training examples: 34,987
Training data: B1K StageGuard single-stage compact-output data, data_improve_V1_60t_task69_removed_delayed2_next_subtask
Augmentation: delayed-switch examples for two frames after a boundary
Task 69: removed entirely

The target is the compact StageGuard decision format rather than a long chain-of-thought trace.

Evaluation

Closed-loop evaluation used 10 unseen demonstrations for each of six B1K tasks (task-0000, task-0001, task-0016, task-0077, task-0087, and task-0092), excluding task 69. The controller accepted a switch immediately (switch-confirm-frames=1). There were 509 ground-truth transitions across 60 trajectories.

Table with columns: Metric, Strict, ±1 frame, ±2 frames, ±3 frames
Metric	Strict	±1 frame	±2 frames	±3 frames
Switch recall	7.07%	20.83%	27.90%	33.40%
Switch precision	—	—	—	47.93%
Switch F1	—	—	—	39.37%

Additional closed-loop results:

Predicted state advances: 459 / 509 ground-truth transitions
Trajectories reaching the final subtask: 49 / 60 (81.7%)
Continue accuracy at ±3: 98.76%
Balanced accuracy at ±3: 66.08%

Frame accuracy is dominated by continue examples, so switch recall, switch F1, and trajectory progression are more informative metrics.

Loading

The checkpoint includes model weights and model configuration. Use the processor/tokenizer from the base model:

python
from transformers import AutoModelForImageTextToText, AutoProcessor

repo_id = "Hoshipu/stageguard-qwen3.5-0.8b-b1k-task69-removed-fullparam-45k"

processor = AutoProcessor.from_pretrained("Qwen/Qwen3.5-0.8B-Base")
model = AutoModelForImageTextToText.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
)

Inputs must follow the StageGuard training prompt schema, including the active subtask and completed-subtask history. Updating that state after an accepted switch is required for closed-loop evaluation.

Limitations

This is a research checkpoint for StageGuard/B1K transition prediction, not a general robot-control policy.
It was evaluated only on the six listed unseen B1K tasks in this experiment.
Switch timing remains imperfect: ±3-frame switch recall is 33.40%.
The repository does not bundle the StageGuard evaluation code or the base model processor files.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Hoshipu

Model Tree

Base

Qwen/Qwen3.5-0.8B-Base

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality