# No code snippets available yet for this library.
# To use this model, check the repository files and the library's documentation.
# Want to help? PRs adding snippets are welcome at:
# https://github.com/huggingface/huggingface.jsThinkMorph-SSL-Base
This is the intermediate Stage 1 model of the PRIMER recipe. It is the EMA weights at step 5,500 of the pretext stage, trained from BAGEL-7B-MoT on Pretext-270K. The final model, ThinkMorph-SSL-Instruct, starts its Stage 2 training from these weights.
This model has only had pretext training, with no question-answer supervision. Its exercises take a corrupted input image and output a description, a restored image and a second description. It is not meant to answer questions directly. The paper's appendix reports roughly 40% on V* and MMMU for this checkpoint. For question answering, use ThinkMorph-SSL-Instruct.
Files
| File | Bytes | SHA256 |
|---|---|---|
ema.safetensors |
29,351,872,574 | f1dd6a3fb88c9145ac8c3d69fd60877cb359ec3996f3a8218580f621e47943ae |
ae.safetensors |
167,664,678 | e49d48aa3474df744ad718d11736d87f38d52bf7ca57dee6b8417978bfbfc9ad |
config.json |
5,175 | 4bc85a4daaef591f0b539216d09a1ee2294a1f3a3bda9f95885c13ed191ae701 |
llm_config.json |
2,136 | cb884fc65d117b8e5ec3856090224cd739c511e01c5c1de488cede5a54c8e994 |
vit_config.json |
1,830 | 78664013bf6778ad4ecaebe78c153ef7fd46f7638fa78c0df1cb17d5bd7df0ab |
generation_config.json |
243 | 3a8f9087e486054c8a4a08dae2e5a3ba62e23da212b5b8c08bc42cb983c3459f |
vocab.json |
3,383,407 | 87a257b04b17642a0688c98cd1df89c398bda4fee532d6f88b38a659ecb4ac8d |
merges.txt |
1,671,853 | 8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5 |
added_tokens.json |
630 | a0582f48ed8e795fda5b7b8eb2d98b492842dc229e78b66793bf545ce7ec52c8 |
special_tokens_map.json |
613 | 76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd |
tokenizer_config.json |
4,864 | b7af0d4c9b917db05150006b466cfa2e8424b3915ff5fe2023a6f512a5e16dd5 |
All files except ema.safetensors are byte-identical to the ThinkMorph-SSL-Instruct release. The
checkpoint's own config and tokenizer files were compared with those of ThinkMorph-SSL-Instruct and
are the same. ema.safetensors uses the same format as ThinkMorph-SSL-Instruct: 1,451 bfloat16 tensors
with the same names and shapes. The unused 26th SigLIP layer is left out.
How the file was made
- The source is the FSDP2 checkpoint of run
0503_stage0_270k_32gpuat step 5,500 (32.distcpshards, 117,699,388,909 bytes). It holds both the regular weights and the EMA weights, all 1,467 tensors of each, in fp32. scripts/merge_checkpoint.py --checkpoint-type emafrom the PRIMER code release merged the EMA state into a single fp32model.safetensors. All 1,467 weights came from the EMA state; none were filled in from a base model.- The 1,451 tensors that ThinkMorph-SSL-Instruct ships were cast once to bfloat16, with no other
change, and saved as
ema.safetensors. A spot check confirmed that the cast values match exactly. The VAE weights are identical to ThinkMorph-SSL-Instruct's, as expected, because the VAE is frozen in both stages.
Training
| Item | Value |
|---|---|
| Run | 0503_stage0_270k_32gpu |
| Initialization | BAGEL-7B-MoT (lmms-lab/BAGEL-7B-MoT-ver.LE layout and VAE) |
| Data | Pretext-270K, 270,896 train rows; the val files are not used |
| GPUs | 32 (4 nodes × 8) |
| Learning rate | 2.83e-5, warmup 1,000 |
| Configured horizon | 10,000 steps; this release is the step-5,500 EMA save |
| EMA | decay 0.9998 from step 1,000, fp32 shadow |
| Losses | text cross-entropy and image flow matching, weights 1 and 1; VAE frozen |
Loading
Loading works the same way as for ThinkMorph-SSL-Instruct. With the official BAGEL inference code or
the PRIMER code release, point model_path at a local copy. The loader reads
ema.safetensors.
License
The weights are released under Apache-2.0, following BAGEL-7B-MoT. Pretext-270K has its own per-source terms; see its card.
- Downloads last month
- 1