Note Eschatia Labs research directory: 3 model families, public models, GGUF builds, datasets and collections.

mertkayacs/Karar-4B

Text Classification • 5B • Updated 3 days ago • 214 • 2

Note Turkish 4B decision model: 96.8% vs Kev-4B 87.1% on held-out decisions from the training data pipeline.

mertkayacs/Deem-4B

Text Classification • 5B • Updated 3 days ago • 215 • 1

Note English 4B decision model: 94.7% vs Kev-4B 84.7% on held-out decisions from the training data pipeline.

mertkayacs/Wahler-4B

Text Classification • 5B • Updated 3 days ago • 265 • 2

Note German 4B decision model: 92.0% vs Kev-4B 81.1% on held-out decisions from the training data pipeline.

mertkayacs/Tholos-2B

Text Generation • 3B • Updated 3 days ago • 910

Note Local tool-use model. Tholos-Bench: 137/160 vs MiniCPM5-2B 112/160; Q4_K_M, llama.cpp JSON schema, Kaggle T4.

mertkayacs/xdfdet

Video Classification • Updated 3 days ago • 718 • 1

Note Eight EfficientNet-B4 detectors with Grad-CAM. Best released AUC: 0.8981 on FaceForensics++; splits differ. Non-commercial weights.

Note Try classification and decisions with probabilities in 3 languages: English, Turkish and German. Use synthetic examples.

mertkayacs/Karar-4B-GGUF

Text Classification • 4B • Updated 3 days ago • 649

Note Turkish decision model for the JevAlt server. Q4_K_M: 2.71 GB file, 3.03 GB peak RAM at 4k context.

mertkayacs/Deem-4B-GGUF

Text Classification • 4B • Updated 3 days ago • 623

Note English decision model for the JevAlt server. Q4_K_M: 2.71 GB file, 3.04 GB peak RAM at 4k context.

mertkayacs/Wahler-4B-GGUF

Text Classification • 4B • Updated 3 days ago • 629

Note German decision model for the JevAlt server. Q4_K_M: 2.71 GB file, 3.03 GB peak RAM at 4k context.

mertkayacs/Tholos-2B-GGUF

Text Generation • 3B • Updated 3 days ago • 368 • 1

Note Local tool use with Ollama or llama.cpp. Q4_K_M file: 1.56 GB; CPU llama.cpp JSON schema: 134/160 tasks.

Note Decision tests: 1,277 English, 1,019 Turkish and 820 German rows, plus 390 live requests. Tests share the training pipeline.

Note 263 executed tool-use trajectories: 237 train, 26 validation. Public sample; most validation runs were used to train Tholos-2B.

Note Probability-labeled training rows in 3 languages: English, Turkish and German. Train, calibration and validation; per-row licenses.

Note Village simulation clips of 3 decision models. Includes a 103-second problem film. Demonstrations, no benchmark scores.