Open research, local inference
Eschatia Labs is an international open-source research lab. We build small language models for decisions and local AI agents, and study explainable deepfake detection.
01 / Model families
Text classification & decisions
Route a support ticket, apply a policy, or choose an action. Three 4B language models return decisions with probabilities through the Jev API.
Held-out tests share the training data pipeline. These scores measure the target decision task. Q4_K_M uses about 3 GB RAM at 4k context with the JevAlt server. Refit calibration on your own data and keep authorization outside the model.
Local AI agents & tool use
A 2B model that chooses tools for tables, notes and tasks on your own machine.
137/160 tasks
Tholos-Bench, Q4_K_M, llama.cpp JSON schema, Kaggle T4. MiniCPM5-2B: 112/160. Only Tholos-2B was fine-tuned for this task format.
Q4_K_M file: 1.56 GB. Use Ollama or llama.cpp. Validate tool calls and retain approvals for consequential actions.
Explainable deepfake detection
Eight EfficientNet-B4 detectors with Grad-CAM maps showing which facial regions influence a prediction.
0.8981AUC on FaceForensics++
Best released checkpoint, aug-cutout-black. Baseline: 0.8684. Each checkpoint used its own random split. On 398 unseen DFDC videos, AUC drops to 0.60 to 0.66.
Weights are CC BY-NC 4.0 under FaceForensics++ terms; code is MIT. A score alone cannot establish a video’s authenticity.
02 / Evidence and training
Cards include sources, split details, licenses and limits.
Decision tests in 3 languages and 390 recorded live requests. Read the split and overlap notes.
263 executed tool-use trajectories. A public sample of a larger training pool.
Training, calibration and validation rows in English, Turkish and German. Check each record’s license.
Recorded village simulations and a 103-second problem film. These clips demonstrate model behavior.
03 / Research collections