---
license: other
library_name: transformers
tags:
  - bert
  - tensorized
  - pytorch
  - rte
  - research-artifact
---

# colaBertFt checkpoint archive

This repository is a preservation archive for checkpoints produced by the
[`colaBertFt`](https://github.com/iAmALai/colaBertFt) research code. The models
use a BERT-large-shaped architecture with tensorized attention and feed-forward
layers. One pair uses low-rank activations throughout; the other restricts the
low-rank activation behavior according to `only_lr_act` in its config.

This is an archival release, not a polished inference package. Loading the
exports may require the custom modeling code in the source repository.

## Contents

| Remote path | Description |
|---|---|
| `bak/chkpt_not_onlylr.pt` | Epoch-2 PyTorch training snapshot with model, optimizer, gradient-scaler, and epoch state. |
| `bak/chkpt_onlylr.pt` | Epoch-2 `only_lr_act` training snapshot with model, optimizer, gradient-scaler, and epoch state. |
| `checkpoints/cola_bert_llama_mlp_large_0.7-004/` | FP16 model state dict and architecture config corresponding exactly to the model tensors in `bak/chkpt_not_onlylr.pt`. |
| `checkpoints/cola_bert_llama_mlp_large_0.7_only_lr-001/` | FP16 model state dict and architecture config corresponding exactly to the model tensors in `bak/chkpt_onlylr.pt`. |
| `model_checkpoints/rte_cola_bert_llama_mlp_large_0.7_only_lr-001/` | PyTorch Lightning RTE checkpoint from epoch 1/global step 68. Its filename records validation accuracy 0.7329; this archive did not independently reproduce that evaluation. |
| `model_checkpoints/colaBert_rte_mlpId_0.7329_2e2e91ae/` | Converted FP16 safetensors RTE export, config, and the original generated model card. |

Exact SHA-256 checksums are stored in `ARTIFACT_CHECKSUMS.sha256` here and in
the source repository.

## Provenance and training

- Base model named by the training configuration: `bert-large-uncased`.
- Downstream task: GLUE RTE.
- RTE Lightning checkpoint metadata: 2 labels, 2,490 training samples,
  learning rate 2e-5, Adam epsilon 1e-8, warmup proportion 0.1, weight decay
  0.01, and 117 planned training steps.
- Archive source files were created in August and September 2025 and archived
  in July 2026.

The two 2.4 GB training snapshots contain optimizer state and are retained for
exact training-resume preservation. Their 615 model tensors match their
corresponding smaller `model.pt` exports exactly.

## Loading and safety

The `.pt` and `.ckpt` files use Python/PyTorch serialization. Only load them
when you trust this archive, and prefer a current PyTorch release with
`weights_only=True` where compatible. The `.safetensors` export avoids pickle
deserialization but still requires the matching custom architecture.

These artifacts are research outputs and have not been audited for production
use, bias, robustness, or generalization beyond the recorded RTE experiment.

## Licensing

The upstream `bert-large-uncased` model is distributed under Apache-2.0. The
source project does not currently declare a license of its own, so this archive
does not make an additional licensing grant for the custom code or derived
artifacts. Users are responsible for checking the applicable upstream terms.