Instructions to use zimplex/hyprag-attnres-checkpoints with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use zimplex/hyprag-attnres-checkpoints with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("zimplex/hyprag-attnres-checkpoints", device_map="auto") - Notebooks
- Google Colab
- Kaggle
HyprAG AttnRes checkpoints
This repository collects the model-only checkpoints requested for the HyprAG
LResNet/TangentAttnRes comparison. Each subdirectory is independently loadable
with the matching hyprag-attnres code. The repository root is not itself a
single model checkpoint.
Checkpoints
| Hub subdirectory | Stage | Architecture | Width | Training point | W&B | Notes |
|---|---|---|---|---|---|---|
stage1/h1512-tangent-native-bf16-d9ogtc44-epoch40 |
1 | TangentAttnRes | 1512 | epoch 40 | d9ogtc44 |
Completed and validated; collaborator-native BF16 residual math |
stage1/h756-tangent-fp32-hardened-iiyrxg14-epoch40 |
1 | TangentAttnRes | 756 | epoch 40 | iiyrxg14 |
Valid epoch-40 export; this older implementation used method-local FP32 for TangentAttnRes log/exp, scoring, and reduction |
stage1/h756-lresnet-knbng6k3-epoch40 |
1 | LResNet | 756 | epoch 40 | knbng6k3 |
Epoch 40 came from the precursor run; d96fgjhn resumed this lineage around epoch 78 |
stage1/h1512-lresnet-e2jdevs4-last-epoch28-flatline |
1 | LResNet | 1512 | last complete epoch 28 | e2jdevs4 |
Partial flatline checkpoint, not epoch 40. Manually cancelled at epoch-equivalent 28.975/40 because loss stayed near 7.40; the run was finite and did not crash |
stage2/h756-lresnet-final |
2 | LResNet | 756 | final | stage2-bertbase-h756-e159-s873823-retry5-ls10-lr1e5-full-20260831T071711Z |
Completed and artifact-validated; initialized from the completed 160-epoch LResNet Stage-1 lineage, not its epoch-40 snapshot |
Checkpoint directory names in the original training outputs were zero-based:
epoch_39_model means 40 completed epochs, while epoch_27/model means 28
completed epochs.
Download one checkpoint
For example:
hf download zimplex/hyprag-attnres-checkpoints \
--include 'stage1/h1512-tangent-native-bf16-d9ogtc44-epoch40/*' \
--local-dir ./hyprag-checkpoints
Use the selected downloaded subdirectory as model_name and
tokenizer_name in the matching Stage-2 configuration. These architectures
depend on the custom modeling implementation in
gdmurphy/hyprag-attnres, so a
stock Transformers AutoModel checkout may not reproduce them by itself.
See manifest.json for exact provenance and SHA-256 hashes of every weight
file.