Spaces:
Running
Running
| title: README | |
| sdk: static | |
| pinned: true | |
| tags: | |
| - authentrics | |
| - machine-unlearning | |
| - model-optimization | |
| - checkpoint-analysis | |
| - catastrophic-forgetting | |
| - backdoor-detection | |
| - interpretability | |
| emoji: 🔬 | |
| colorFrom: blue | |
| colorTo: gray | |
| short_description: We are Authentrics.ai | |
| [](https://app.authentrics.ai/docs) | |
| [](https://app.authentrics.ai/) | |
| [](https://app.authentrics.ai/) | |
| # Authentrics.ai | |
| **Find what changed. Find when it changed. Fix it without retraining.** | |
| Training is one of the most expensive and consequential parts of building a model. Once a run is | |
| complete, regressions, memorization, interference, or unwanted behavior can leave teams with a | |
| familiar answer: re-curate the dataset, change the training setup, and run it again. | |
| Authentrics provides another path — analyze how the model changed, trace where problems emerged, | |
| and correct the resulting model without a full retrain. | |
| ## What You Can Do With Authentrics | |
| - **Understand your model.** Establish a baseline across model behavior and internal structure, | |
| detect meaningful changes, and localize which layers are most associated with them. | |
| - **Remove a known cause (machine unlearning).** When unwanted behavior traces back to specific training | |
| session(s), remove that effect directly from the checkpoint. | |
| - **Model optimization.** Recover accuracy, correct a regression, or suppress unwanted behavior by | |
| optimizing the checkpoint directly against an objective — without backprop or retraining. | |
| Everything runs locally on your own hardware; only project metadata leaves your machine, never your weights. | |
| | Operation | What it's for | | |
| | --- | --- | | |
| | `static_analysis` | audit parameter drift between pairs of sequential checkpoints | | |
| | `activation_analysis` | catch behavioral drift in intermediate activations | | |
| | `correlation_analysis` | see how strongly each layer influences a reference (usually output) layer | | |
| | `exclude_training` | compliance / unlearning: remove the effect of specific training session(s) (e.g. a sensitive/bad data slice) from the latest checkpoint without retraining from scratch | | |
| | `ztom_analysis` | optimize a model against a Callable[[output], float] loss without backprop/retraining | | |
| ## Demos | |
| Each demo starts with a familiar machine-learning problem and shows how Authentrics approaches it | |
| through direct analysis and modification of the model itself. | |
| ### Find Which Training Step Introduced a Backdoor — Then Remove It | |
| *Data poisoning / backdoor detection / machine unlearning* | |
| A LeNet-5 MNIST classifier is trained with one poisoned epoch that introduces a BadNets trigger. Clean accuracy remains healthy, while triggered inputs are silently misclassified. | |
| Authentrics analyzes how the model changed through training, identifies where the unwanted behavior emerged, and removes that learned influence without retraining from scratch. | |
| **Demonstrates:** training dynamics, backdoor detection, machine unlearning | |
| - [Source code (GitHub)](https://github.com/Authentrics-ai/trojai-authentrics-demo) | |
| - [Model on Hugging Face](https://huggingface.co/authentrics/mnist-badnets-poisoned) | |
| - [Corrected model](https://huggingface.co/authentrics/mnist-badnets-repaired) | |
| --- | |
| ### Recover a Capability Lost During Fine-Tuning | |
| *Catastrophic forgetting / task interference / model optimization* | |
| Fine-tuning Gemma 3 1B on a new task quietly degrades a capability the model already had. | |
| Instead of retraining or repeating fine-tuning, Authentrics works directly with the resulting model to recover the lost behavior. | |
| **Demonstrates:** catastrophic forgetting, regression recovery, model optimization | |
| - [Source code (GitHub)](https://github.com/Authentrics-ai/authentrics-analysis-examples/blob/main/hf_legal_understanding.ipynb) | |
| - [Model on Hugging Face](https://huggingface.co/authentrics/ztom-gemma3-1b-recovery) | |
| --- | |
| ### Remove Memorized PII Without Retraining the Model | |
| *Memorization / privacy / machine unlearning* | |
| A medical chatbot fine-tuned on data containing personal information memorizes that information. | |
| Authentrics identifies where the unwanted learning was introduced and removes its influence directly from the model without a full retrain. | |
| **Demonstrates:** memorization analysis, training attribution, machine unlearning | |
| - [Source code (GitHub)](https://github.com/Authentrics-ai/authentrics-analysis-examples/blob/main/hf_medical_chatbot.ipynb) | |
| - [Model on Hugging Face](https://huggingface.co/authentrics/pii-removal-medical-llama32) | |
| --- | |
| ### Understand What Changed Inside Your Model | |
| #### See Which Experts a Mixture-of-Experts Model Actually Uses | |
| *Interpretability / Mixture-of-Experts routing* | |
| Which experts does Mixtral-8x7B actually use, and how does that internal routing behavior vary across the model? | |
| Authentrics analyzes the model's internal structure and behavior to expose routing patterns and layer-level influence. | |
| **Demonstrates:** model interpretability, MoE routing, internal model analysis | |
| - [Source code (GitHub)](https://github.com/Authentrics-ai/authentrics-model-analysis-experiments#mistralmixtral-moe-expert-activation) | |
| - [Model on Hugging Face](https://huggingface.co/authentrics/mixtral-expert-routing) | |
| #### Find Where a Model Changed Across Its Training Run | |
| *Training dynamics / model drift* | |
| Across the NVIDIA Nemotron Cascade 8B training lineage, what changed inside the model — where, when, and how much? | |
| Authentrics analyzes the evolution of the model across training to surface where meaningful structural and behavioral changes occurred. | |
| **Demonstrates:** training dynamics, model drift, model evolution | |
| - [Source code (GitHub)](https://github.com/Authentrics-ai/authentrics-model-analysis-experiments) | |
| - [Model on Hugging Face](https://huggingface.co/authentrics/nemotron-training-dynamics) | |
| ## Try It on Your Model | |
| ```bash | |
| pip install authentrics # Linux x86_64, Python 3.11–3.13 | |
| authrx init # paste API key (stored at ~/.local/state/authentrics/api_key) | |
| # or, for CI / non-interactive: | |
| export AUTHRX_API_KEY=<your_api_key> | |
| ``` | |
| Generate an API key and read the full docs at [https://app.authentrics.ai/](https://app.authentrics.ai/). | |
| ## Links | |
| - **App & API keys:** [https://app.authentrics.ai/](https://app.authentrics.ai/) | |
| - **Docs & API reference:** [https://app.authentrics.ai/docs](https://app.authentrics.ai/docs) | |
| - **Examples & user guide:** [https://github.com/Authentrics-ai/authentrics-analysis-examples](https://github.com/Authentrics-ai/authentrics-analysis-examples) | |
| - **Contact:** [info@authentrics.ai](mailto:info@authentrics.ai) | |