| --- |
| tags: |
| - tensorbert |
| - bert |
| - glue |
| - pytorch |
| --- |
| |
| # TensorBERT-FT artifacts |
|
|
| Training checkpoints and evaluation outputs produced by the |
| [TensorBERT-FT](https://github.com/alvin-zyl/TensorBERT-FT) codebase. |
|
|
| ## Local artifact inventory |
|
|
| Inventory taken on 2026-07-30. Sizes are binary units (GiB/MiB). |
|
|
| | Local path | Size | Contents | |
| |---|---:|---| |
| | `tensor-/results/glue/` | 32.3 GiB | 35 fine-tuned GLUE checkpoints plus configs, metrics, and logs | |
| | `weights/` | 13.2 GiB | 6 pretrained/intermediate `.pt` checkpoints | |
| | `glue_data/` | 2.9 GiB | Local GLUE data and preprocessing caches; intentionally not uploaded | |
| | Remaining source tree | < 3 MiB | Code, notebooks, and metadata | |
| | **Total local tree** | **48.1 GiB** | 396 regular files | |
|
|
| By file type, the large artifacts are approximately 32.3 GiB in 35 `.bin` |
| files, 13.2 GiB in 6 `.pt` files, and 1.4 GiB in 43 extensionless files. |
| There are 46 files larger than 100 MiB, totaling about 48.0 GiB. |
|
|
| The local directory's measured size was 49 GiB as rounded by `du -h`, both |
| for allocated and apparent size. It was not 76 GB at inventory time. |
|
|
| ## Checkpoint groups |
|
|
| The 35 fine-tuned result checkpoints are each about 944 MiB: |
|
|
| | Group | Checkpoints | Approximate size | |
| |---|---:|---:| |
| | CoLA | 2 | 1.8 GiB | |
| | MRPC | 10 | 9.2 GiB | |
| | RTE `0_7-004` sweep | 10 | 9.2 GiB | |
| | RTE `0_7_only_lr-001` sweep | 10 | 9.2 GiB | |
| | Other RTE | 2 | 1.8 GiB | |
| | WNLI | 1 | 0.9 GiB | |
|
|
| The six files under `weights/` are: |
|
|
| | Path | Approximate size | |
| |---|---:| |
| | `weights/bert_large/model.pt` | 3.2 GiB | |
| | `weights/bert_large/ckpt_8038.pt` | 3.2 GiB | |
| | `weights/cola_bert_llama_mlp_large_0.7-004/model.pt` | 2.4 GiB | |
| | `weights/cola_bert_llama_mlp_large_0.7_only_lr-001/model.pt` | 2.4 GiB | |
| | `weights/tensor_bert/model.pt` | 1.2 GiB | |
| | `weights/tensor_bert/ckpt_8038.pt` | 1.2 GiB | |
|
|
| ## Curated evaluation checkpoints |
|
|
| The curated upload profile selects the best observed validation result in |
| each available task/group: |
|
|
| | Task | Local run | Validation metric | |
| |---|---|---:| |
| | CoLA | `COLA-0_lr_only` | MCC 0.4519 | |
| | MRPC | `MRPC-3_32_8` | accuracy/F1 mean 0.8774 | |
| | RTE | `RTE-0_7_only_lr-001_6_32_8` | accuracy 0.6968 | |
| | WNLI | `WNLI` | accuracy 0.5634 | |
|
|
| These values are copied from each run's `results.txt`; they are not an |
| independent reproduction. |
|
|
| ## Duplicate note |
|
|
| SHA-256 inventory found that `results/glue/RTE/pytorch_model.bin` and |
| `results/glue/RTE-0_7-004_3_32_8/pytorch_model.bin` are byte-identical. The |
| full-results upload omits the former duplicate directory. Other large files |
| had distinct SHA-256 hashes. |
|
|
| ## Scope and reuse |
|
|
| The GLUE dataset files and preprocessing caches are not included. Obtain the |
| original datasets from their authoritative distributors and follow their |
| respective licenses. The license/provenance of the checkpoint weights should |
| be confirmed by the uploader before representing these artifacts as freely |
| redistributable. |
|
|
|
|
