BAGEL navigation points — 3,000-step checkpoint
This is the compact-point Spatial Navigation checkpoint from the Primitive
Imagination project. It fine-tunes BAGEL's understanding path to emit an
ordered sequence of normalized <point>x,y</point> tokens describing a route.
It does not generate or re-observe an intermediate image.
Result
| Evaluation | Result |
|---|---|
| Released ThinkMorph/VSP navigation test, 600 examples (100 each at 3×3–8×8) | 453/600 (75.5%) strict route success |
Strict success requires the emitted ordered points to execute from start to goal without crossing a hole or leaving the grid. This number uses corrected native-resolution BAGEL inference, not the invalid early vLLM preprocessing.
Training
- Base:
ByteDance-Seed/BAGEL-7B-MoT - Data: 6,000 released Spatial Navigation training problems, sizes 3×3–6×6
- Objective: selective CE on point spans; no raster-thought image target
- Updates: 3,000 on four GPUs
- Adaptation: full understanding-branch fine-tuning, not LoRA
- Generation expert: unchanged
- Coordinates: normalized
[0,1000]
Files and loading
understanding.safetensors contains the 755 tensors saved by the fine-tuned
understanding path. Obtain the upstream BAGEL checkpoint, then export a merged
checkpoint with this repository:
- Size: 16,105,547,536 bytes
- SHA-256:
26893d02cf7ba3fd206bcb58e43af37e09dd72a9c28f5c4c228f3995311be417
git clone --recurse-submodules https://github.com/oshapio/primitive-imagination-lab
cd primitive-imagination-lab/experiments/bagel-navigation
python tools/export_checkpoint.py \
--base /path/to/BAGEL-7B-MoT \
--checkpoint /path/to/downloaded/navigation-points-delta \
--out /path/to/navigation-points-merged
The downloaded delta directory must contain the file under the exact name
understanding.safetensors. The export retains the upstream BAGEL files and
replaces only the trained tensors. Native inference is implemented by
src/model_loader.py; evaluation is evaluate.py.
Important limitations
- This is a research artifact for a narrow synthetic navigation task, not a general navigation system.
- The released VSP source corpus and evaluation set are not fully independent: the project audit found 63 overlapping boards / 61 source-training records. Do not present 75.5% as clean unseen-layout generalization.
- Selective point-only supervision has a format/stopping mismatch with the original boxed-answer prompt. Raw generations and parse failures remain part of the evaluation record.
- The checkpoint does not establish that point traces are causally necessary; mask/swap/truncation controls are still planned.
Evidence
- Code and complete experiment map
- Raw checkpoint comparison and overlays
- Exact training targets
- Native ThinkMorph visual-SFT comparison
Upstream BAGEL and ThinkMorph licensing and attribution continue to apply.