Open artifacts
What the lab builds
A measurement lab that only measures is hard to check. These are the artifacts the lab publishes in the open — models, a training dataset, the harness and the evidence archive — each under a permissive license, each solving a problem the lab hit in its own work. Performance figures are not stated here: measured numbers live in the claim registry.
- Model License: Apache-2.0 Base: deepvk/RuModernBERT-small
strizh-ru-retriever
Problem it solves
Self-hosted Russian RAG needs a first-stage retriever small enough to run on CPU next to everything else — and one that does not require query/document prefixes, which every wrapper forgets to add.
What it is
A 24.4M-parameter, 4-layer sentence encoder layer-pruned and re-trained from RuModernBERT-small. One 384-dimensional embedding per text, no prefixes, architecture accepts up to 8192 tokens (contrastive training used 256-token sequences).
- Model License: Apache-2.0 Base: t-tech/T-lite-it-2.1 (Qwen3-8B)
agmind-rag-splitter-ru
Problem it solves
Fixed-size chunking cuts Russian documents mid-thought and splits tables and code blocks in half. Retrieval then returns fragments that no longer answer anything on their own.
What it is
A LoRA fine-tune that takes a document pre-split into numbered units and returns boundary indices plus a topic, as JSON — chunks are assembled host-side. v2 raised the dataset to 28k examples across six source types and added a held-out reference set of 140 out-of-domain legal documents labeled by consensus of two independent models with arbitration.
- Dataset License: Apache-2.0
agmind-rag-splitter-ru-data
Problem it solves
Nobody publishes the training data behind a Russian chunking model, so nobody can check what it learned to consider a boundary.
What it is
The training corpus behind the splitter, published so the labels can be inspected and the model retrained by anyone.
- Model License: Apache-2.0 Base: Qwen/Qwen3-1.7B
qmd-query-expansion-ru
Problem it solves
The stock query-expansion model shipped with the qmd local search engine is trained on English only: it rewrites a Russian query into an English hallucination template and its output format falls apart (upstream issues #774 and #454). qmd's BM25 also uses a Porter stemmer that does not understand Russian morphology, so the expansion itself has to cover word forms.
What it is
A drop-in Russian replacement trained with the upstream finetune recipe, producing both semantic and morphological expansions so lexical search finds the inflected forms.
- Harness License: Apache-2.0
agmind-bench
Problem it solves
A published number is worth nothing if the instrument that produced it is private: nobody can check what counted as a failure, or reproduce the cell.
What it is
The measurement harness itself: the runner, the frozen workload corpora, the six quality gates and their unit tests, plus the DuckDB queries that derive every published value.
- Evidence archive License: CC BY 4.0
agmind-lab
Problem it solves
Summary tables can hide anything. Without per-request records an outsider cannot tell a clean result from a curated one.
What it is
Every published run as a sealed evidence bundle: manifest, per-request records including failures, per-request gate outcomes, harness aggregates, server log and checksums — plus the catalog that ties bundles to claims.
Numbers describing an artifact (parameter count, embedding size, dataset size) are properties of the artifact itself, verifiable on its repository page. Anything about how fast or how well a system runs is measured, and lives with its evidence in the claim registry.