AGmind-ML - GitLite

KO | EN

gitlite — search

#cross-platform #gpt #ruby #nextjs #csharp #pytorch #cpp #css #web #rust

AGmind-ML

(★ 21)

Locally fine-tuned Russian RAG models: document splitter, query expansion, retrieval embedder. Teacher distillation → GGUF → llama.cpp on AMD Vulkan. Commercial-OK licenses.

File Explorer

ZIP Download

- - generate.py
  - README.md
  - sample_holdout.jsonl
  - sample_train.jsonl
- - methodology.md
  - model_card.md
- - eval_gguf.py
  - eval_gguf_2026-06-26.log
  - eval_hf.py
  - eval_hf_n1500_2026-06-26.log
  - eval_hf_n300_2026-06-26.log
  - results.md
- - demo.py
  - segmenter.py
  - wrapper_openai.py
- - test_segmenter.py
- - config.md
  - merge_lora.py
  - patch_tokenizer_hash.py
  - train.py
- README.md
- requirements.txt
- bench_qmd.py
- gen_qmd.py
- README.md
- train_qmd.py
- - analyze_prompt_tokens.py
  - batch_thr.py
  - bench_v2.py
  - context_len_check.py
  - coresident_bench.py
  - corpus_ru.json
  - cross_lingual.py
  - dev_clean_ru.py
  - dev_one.py
  - en_nanobeir.py
  - extended_bench.py
  - gguf_eval.py
  - index_load_check.py
  - loadtest_v2.py
  - mixed_bench.py
  - pipeline_bench.py
  - pipeline_bench_v1.py
  - prompt_identity_check.py
  - prompt_tokens_check.py
  - README.md
  - results-coresident-strix-2026-07-27.txt
  - results-embed-curves-strix-2026-07-27.txt
  - results-flags-control-strix-2026-07-28.txt
  - results-prompt-identity-strix-2026-07-28.txt
  - results-prompt-tokens-strix-2026-07-28.txt
  - results-rag-final-strix-2026-07-27.txt
  - results-rag-pipeline-strix-2026-07-27.txt
  - results-ru-filtered-2026-07-23.txt
  - results-speed-strix-2026-07-26.txt
  - texts_long_ru.json
  - texts_short_ru.json
- data_v2.py
- dev_one.py
- inject_epsilon.py
- mine_bge.py
- README.md
- train_student.py
- train_v2.py
- train_v2i2.py
- warmstart_sweep.py
.gitignore
CITATION.cff
LICENSE
README.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

**Locally fine-tuned Russian models for the self-hosted [AGmind](https://github.com/botAGI/AGmind) RAG stack.**
Teacher distillation, quantization to GGUF, inference on **AMD (Vulkan, no CUDA)** via
`llama.cpp`. Commercial-OK licenses, fully local. Trained on a single consumer GPU
(RTX 5090), deployed on AMD Strix Halo. Three models shipped so far — a document
splitter, query expansion, and a retrieval embedder.

![License](https://img.shields.io/badge/license-Apache--2.0-blue.svg)
![Serve](https://img.shields.io/badge/inference-AMD%20Vulkan-red)
![Lang](https://img.shields.io/badge/lang-RU-informational)

## Models

| Project | What it does | Base | Key numbers | Status |
|---|---|---|---|---|
| [**agmind-rag-splitter-ru**](agmind-rag-splitter-ru/) · [🤗](https://huggingface.co/AGmind/agmind-rag-splitter-ru) | Context-aware document splitter for RAG: semantic chunks, tables and code kept whole, lossless reconstruction | `t-tech/T-lite-it-2.1` (Qwen3-8B) | 100% valid JSON, boundary-F1@±1 0.821 | ✅ on HF |
| [**qmd-query-expansion-ru**](qmd-query-expansion-ru/) · [🤗](https://huggingface.co/AGmind/qmd-query-expansion-ru) | Search-query expansion for [qmd](https://github.com/tobi/qmd) (hyde/lex/vec, word forms for BM25); fixes the stock model's English hallucinations on Russian | `Qwen/Qwen3-1.7B` | reward-score 94.8; 100% format | ✅ on HF |
| [**strizh-ru-retriever**](strizh-ru-retriever/) · [🤗](https://huggingface.co/AGmind/strizh-ru-retriever) · [GGUF](https://huggingface.co/AGmind/strizh-ru-retriever-GGUF) | RU-first no-prefix dense retriever: 4 layers, distilled by layer-pruning + BGE-M3 hard negatives. A speed/drop-in specialist — beats rubert-tiny2 on Russian, below USER2-small/bge-m3 on quality | `deepvk/RuModernBERT-small` | RU recall@10 0.75 (clean) / EN nDCG@10 0.34; 24M, fast on Strix Vulkan (slower on CPU) | ✅ on HF |
| _next…_ | _(150M embedder, reranker, grounded-QA)_ | | | planned |

## Method

- **Teacher distillation.** A strong model labels the task or defines the target space —
  no manual annotation, with strict validation gates, dedup and retries.
- **One RTX 5090 (32 GB).** Generative models: bf16 LoRA (+rsLoRA), response-only loss.
  Embedders: teacher layer-pruning (warm-start) + contrastive (MNRL) on hard negatives.
- **Cyrillic bases.** Russian-tuned tokenizers are more economical than vanilla ones.
- **Deploy anywhere.** GGUF → `llama.cpp` Vulkan on AMD Strix Halo, no CUDA.

## Layout

```
AGmind-ML/
├── agmind-rag-splitter-ru/   # document splitter
├── qmd-query-expansion-ru/   # query expansion
└── strizh-ru-retriever/      # retrieval embedder
```

Each project has its own `README.md`: task, method, metrics, reproduction.