Downloads · 30 days
4
15% of all-time downloads
wjdghks950/promac-futloc-stage2-stage0init-visfrozen-step200
promac-futloc-stage2-stage0init-visfrozen-step200 is a image-text-to-text model from wjdghks950. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as other.
Public, inference-only Qwen3-VL export of the exact ProMAC FutLoc Stage-2 checkpoint at optimizer globalstep200, trained with Stage-0 predictor initialization and a frozen Qwen vision tower.
Downloads · 30 days
4
15% of all-time downloads
All-time downloads
27
Public
Parameters
4.4B
8.9 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors8.9 GB · 100%
From the Hugging Face model README
Public, inference-only Qwen3-VL export of the exact ProMAC FutLoc Stage-2 checkpoint at optimizer global_step200, trained with Stage-0 predictor initialization and a frozen Qwen vision tower.
On all 3,016 examples of the FutLoc structured-CoT v4 goal-overlay validation split (futloc-v4-val), deterministic evaluation reported:
| Checkpoint | ADE | FDE | L2 First 4 | DTW | DTW % path | Miss @20% |
|---|---|---|---|---|---|---|
checkpoint-best-ade/global_step200 | 123.96 | 119.33 | 115.44 | 87.320 | 12.8 | 46.2 |
The DeepSpeed tag is pinned explicitly and corresponds to training micro-step 800 with gradient accumulation 4. FutLoc coordinates are pixels in the 1280 x 720 frame over 40 waypoints.
import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
model_id = "wjdghks950/promac-futloc-stage2-stage0init-visfrozen-step200"
processor = AutoProcessor.from_pretrained(model_id, fix_mistral_regex=True)
model = Qwen3VLForConditionalGeneration.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
Use the FutLoc structured-CoT v4 goal-only-overlay prompt schema and answer_only assistant format. Set max_new_tokens=1024; 512 tokens truncated trajectories in the full offline evaluation.
This repository contains the Qwen3-VL subtree extracted from the pinned ProMAC ZeRO-3 tag and merged into a standard Hugging Face model. Trajectory inference uses the standard Qwen generate() path. The Stage-0-initialized future-state predictor was used during training but is not required for this inference path and is not included. The auxiliary decoder, SIGReg state, optimizer state, raw ZeRO shards, logs, and datasets are excluded.
The export was merged with Qwen/Qwen3-VL-4B-Instruct revision ebb281ec70b05090aa6165b016eac8ec08e71b17. See provenance.json and SHA256SUMS.
The upstream Qwen model card declares Apache-2.0. No standalone license was established locally for the ProMAC fine-tuned weights, so this repository uses license: other. Ego4D data is distributed through a license-gated workflow; access to these weights does not grant rights to source data or derived imagery.
This checkpoint is an intermediate optimizer step, intentionally pinned rather than resolved through the mutable latest pointer. It was evaluated on the documented FutLoc validation split and may not generalize to other cameras, prompts, environments, or trajectory schemas. Reported metrics are not safety guarantees.