| --- |
| license: mit |
| base_model: Qwen/Qwen3.5-0.8B |
| library_name: pytorch |
| tags: |
| - interpretability |
| - activation-probing |
| - vision-language-models |
| - temporal |
| - qwen |
| --- |
| |
| # TM-NLA Checkpoints |
|
|
| These checkpoints support the TM-NLA public runtime. |
|
|
| TM-NLA is a proof-of-concept temporal activation-verbalization probe for frozen visual-temporal hidden states. It projects video-window activations from `Qwen/Qwen3.5-0.8B` into a continuous language-oriented manifold, then selects semantic readout hypotheses over time with compatibility and specificity annotations. |
|
|
| ## Files |
|
|
| | File | Role | SHA-256 | |
| | --- | --- | --- | |
| | `temporal_probe.pt` | Temporal visual activation -> continuous language-manifold probe | `48AABF662423EA529C985855EA54970A9CE3256EC3F14AFE0169C2BF1A55B217` | |
| | `activation_verbalizer.pt` | Activation -> small set of short candidate readout proposals | `88DDB0267AE535A287755DE5C9E638EB771FFD5C5377051ACFCADF96A98C08C1` | |
| | `text_reconstructor.pt` | Text readout -> activation compatibility verifier/reranker | `C5CEA721A2173DCA884617258BB3F671668CF13EBA07728E11C39C9695B5DE32` | |
|
|
| ## Base Model |
|
|
| The runtime uses frozen `Qwen/Qwen3.5-0.8B` activations. The public demo path does not train or fine-tune Qwen. |
|
|
| ## Intended Use |
|
|
| Use these checkpoints with the TM-NLA GitHub repository to run local terminal-based semantic readouts over video windows: |
|
|
| https://github.com/DavWer/TM-NLA |
|
|
| The checkpoints are intended for interpretability and research exploration. They are not intended as a production captioner, classifier, benchmark system, or broad video-analysis product. |
|
|
| ## Limitations |
|
|
| The continuous activation geometry is the strongest result. TM-NLA shows weak recoverable temporal semantic traces from frozen visual-temporal activations. Natural-language readouts are useful but noisy externalizations, not ground-truth captions. |
|
|
| Readouts can be malformed, wrong, generic, or underspecified. The activation verbalizer remains fragile, and candidate generation plus AR reranking is a practical readout mechanism rather than a solved decoder. The text reconstructor is a compatibility-based verifier/reranker, not an oracle. `specificity_status` is a numerical activation-specificity annotation, not a ground-truth correctness label. |
|
|
| ## Citation |
|
|
| If you publish work building on this proof of concept, cite or link the companion TM-NLA GitHub repository: |
|
|
| https://github.com/DavWer/TM-NLA |
|
|
