README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-VL-4B-Instruct
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- strive
- video
- qwen3-vl
- kubric
- scene-understanding
---

# STRIVE Kubric checkpoints

This repository contains the three merged checkpoints used for the STRIVE
Kubric baseline experiments, plus six intermediate LoRA adapters for Baselines 2 and 3. The corresponding code, commands, evaluation
artifacts, and detailed provenance are in the
[STRIVE reproducibility repository](https://github.com/suryathecreator/strive/blob/main/REPRODUCIBILITY.md).

All three checkpoints are fine-tuned and merged derivatives of
[`Qwen/Qwen3-VL-4B-Instruct`](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct). They retain the base
model's Apache-2.0 license.

## Checkpoints

| Subfolder | Step | Size | Purpose |
|---|---:|---:|---|
| `baseline1-detector-5500` | 5500 | 8.28 GiB | Single-frame object detector used as Baseline 2 stage 1. |
| `baseline2-4000` | 4000 | 8.28 GiB | Streaming staged mapper used as Baseline 2 stage 2. |
| `baseline3-4000` | 4000 | 8.28 GiB | Streaming end-to-end scene parser used as Baseline 3. |

Baseline 2 is staged: use `baseline1-detector-5500` for stage-one detection
and `baseline2-4000` for stage-two mapping. Baseline 3 uses
`baseline3-4000` directly.


## Intermediate LoRA adapters

These PEFT adapters preserve the retained pre-final checkpoints without
duplicating the base model or training-only optimizer state.

| Subfolder | Step | Purpose |
|---|---:|---|
| `baseline2-adapter-3250` | 3250 | Baseline 2 stage-two mapper/tracker |
| `baseline2-adapter-3500` | 3500 | Baseline 2 stage-two mapper/tracker |
| `baseline2-adapter-3750` | 3750 | Baseline 2 stage-two mapper/tracker |
| `baseline3-adapter-3250` | 3250 | Baseline 3 end-to-end scene parser |
| `baseline3-adapter-3500` | 3500 | Baseline 3 end-to-end scene parser |
| `baseline3-adapter-3750` | 3750 | Baseline 3 end-to-end scene parser |

Each adapter targets `Qwen/Qwen3-VL-4B-Instruct` and includes its PEFT
configuration, adapter safetensors, tokenizer, chat template, and processor
configuration. Optimizer shards, scheduler state, RNG state, and other
training-resume artifacts are intentionally excluded.

## Download

Download one checkpoint without fetching the other two:

```python
from huggingface_hub import snapshot_download

repo_id = "suryadv/strive-kubric-checkpoints"
subfolder = "baseline3-4000"
snapshot = snapshot_download(
    repo_id,
    allow_patterns=[f"{subfolder}/*"],
)
model_path = f"{snapshot}/{subfolder}"
print(model_path)
```

The returned `model_path` can be passed to Transformers, vLLM, or the STRIVE
inference launchers as a normal local merged-model directory. Transformers
also accepts the repository ID together with `subfolder=<checkpoint-name>`.

## Integrity and provenance

`CHECKPOINT_MANIFEST.json` records the exact byte size and SHA-256 digest of
every uploaded file. Each subfolder contains the full merged checkpoint:
two safetensors shards, their index, model/generation configuration, tokenizer,
processor configuration, chat template, and vocabulary files. Training-only optimizer state, caches, datasets, and evaluation outputs are
+intentionally not duplicated here. The intermediate adapter folders contain only
+the reusable LoRA and inference metadata described above.

The training and evaluation setup is documented at the reproducibility tag
[`strive-repro-2026-07-18`](https://github.com/suryathecreator/strive/releases/tag/strive-repro-2026-07-18).