Eschatia Labs labs
Menu
EN

A small collective of researchers

Eschatia Labs is an independent, non-profit open-source AI research lab.

Eschatia Labs is a small group of researchers working as a collective. We collect our preprints here, publish our models on Hugging Face and our code as open source, and share all of it with everyone.

Open benchmarks GitHub Hugging Face
11,397downloads of our models and datasets on Hugging Face
10,753downloads of our code packages on npm
7open projects
3open benchmarks

Counts as of

Research

Eschatia Labs · five areas · one method

01 Decision models A decision model reads a situation, picks one of the given answers and gives each answer a probability. Deem-4B works in English, Karar-4B in Turkish and Wähler-4B in German. All three were trained from Intern-Decision-4B.On jevalt-bench, Deem-4B reaches 94.7% accuracy, versus 84.7% for Kev-4B. 02 Open benchmarks A benchmark is a fixed set of test questions with a scoring rule. jevalt-bench has 3,116 written situations in three languages. JevOss checks whether an answer changes when the question is rewritten a little. Tholos-Bench has 160 tasks in a shared workspace of tables, notes and a task board.3 suites · 3,116 situations · 160 workspace tasks 03 Explainable deepfake detection A deepfake is a video in which AI has changed or replaced a face. xdfdet checks each video frame with EfficientNet-B4, an image recognition model. Grad-CAM heat maps show which parts of the face each decision relied on. AUC scores a detector from 0.5 (guessing) to 1 (perfect).On FaceForensics++, EfficientNet-B4 reaches 0.897 AUC with varied and partly blacked-out training images, versus 0.868 without them. 04 Local task assistants A local task assistant does work on your own computer under rules you set. Tholos assistants share tables, notes and a task board. Each action runs, waits for your approval or is blocked.On Tholos-Bench, on a Kaggle T4 graphics card, Tholos-2B passes 137 of 160 tasks, versus 112 for MiniCPM5-2B. 05 Language preservation Eldalambë helps people read and write Quenya and Sindarin, the Elvish languages Tolkien invented, and their script, Tengwar. A small language model checks its answers against a dictionary and grammar rules. The paper is public; the code, data and model files are private.Public preprint · code, data, models and test set private

Projects

Seven projects from Eschatia Labs

Game

Emberwick

A medieval village you play in the browser. Three small decision models, Deem-4B, Karar-4B and Wähler-4B, decide what the villagers do through storms, fires, markets and wolves.

Open project
Developer tool

reevesagents

Runs several AI coding assistants (Claude Code, Codex and 20+ others) side by side in one terminal, with one assistant coordinating the rest.

Open project
Decision models

JevAlt

The three decision models behind Emberwick, for English, Turkish and German. Each answer comes with a probability, and the models run on a laptop with about 3 GB of memory.

Open project
Computer vision

xdfdet

Explainable deepfake detection: eight EfficientNet-B4 models that spot AI-altered faces frame by frame, with Grad-CAM heat maps as evidence. The MSc thesis and the UBMK 2026 paper come from this work.

Open project →
Local assistant

Tholos

Small local assistants that share tables, notes and a task board on your own computer. A rule checks each action first: it runs, waits for your approval or is blocked.

Open project
Benchmark

JevOss

An open test suite for question drift, which is when a model changes its answer after a small rewrite of the question.

Open project
Language

Eldalambë

A small model for Quenya and Sindarin, the Elvish languages Tolkien invented, and their script, Tengwar. The paper is public; the code and data are private.

Open project

Open benchmarks from Eschatia Labs

Test questions stay out of training. Results, failures, and reproduction steps are published.

jevalt-bench has 3,116 written situations. JevOss checks question drift, meaning answer changes caused by small question rewrites. Tholos-Bench has 160 workspace tasks where assistants edit tables, notes, and assignments.

Open the benchmarks

Publications

Eschatia Labs preprints, papers and technical reports

UBMK 2026

Augmentation and Cutout in Deepfake Detection: A Comparative Study of Accuracy, Calibration, and Attention

Compares nine ways to train a deepfake detector: how accurate it is, how well its confidence matches its results, and which parts of the face it relies on. Accepted at UBMK 2026 (IEEE), to appear in IEEE Xplore.

Accepted
Preprint

Language Preservation Using Agentic AI Architectures: A Tool-Grounded Small Language Model on a Constructed-Language Testbed

A small language model for Quenya and Sindarin, the Elvish languages Tolkien invented, with tools that check every answer against a dictionary and grammar rules.

Public
Technical report

JevAlt: Technical report

How Deem-4B, Karar-4B and Wähler-4B were trained and tested, with results, limits and download links.

Public
Technical report

Tholos-2B: Technical report

How Tholos-2B was trained from MiniCPM5-2B, and how it scores on the 160 Tholos-Bench tasks.

Public
MSc thesis

Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques

The thesis behind xdfdet: EfficientNet-B4 detectors tested frame by frame on FaceForensics++, with Grad-CAM heat maps. TED University, 2025. DOI 10.5281/zenodo.18998566.

Zenodo

Approach

One model, one job, one defined test.

  1. IEach model is trained for one job and tested on that job.
  2. IITest questions stay out of the final training run.
  3. IIIResults, known failures and the files to repeat them are published together.
  4. IVThe released JevAlt files run on a laptop processor with about 3 GB of memory.