AI Research Projects

Selected LLM, NLP, retrieval and vision-language research-engineering projects.

This repository collects compact AI research-engineering projects around language-model training, dense retrieval, information extraction and multimodal search. Each project includes a README, reproducibility notes, limitations and excludes raw datasets, checkpoints and generated artifacts.

Featured Projects

Russian LLM Pretraining and SFT

russian_llm_pretraining_and_sft

• Compact LLM training workflow: corpus preparation, tokenizer training, causal LM pretraining, LoRA SFT and conservative evaluation reporting.
• Documented snapshot:
  • 44,432 pretraining train blocks
  • 917 validation blocks
  • 7,840 SFT train examples
  • SFT training loss 1.6505
• This is a small project-local experiment, not a production-ready model-quality claim.

Semantic Retrieval for arXiv Papers

arxiv_semantic_retrieval

• Dense retrieval over arXiv metadata with BGE embeddings and FAISS.
• Documented metrics:
  • 98,213 documents
  • 1,000 evaluation queries
  • MRR@5 0.9164
  • HitRate@5 0.9650

Multi-Task Information Extraction on NEREL

nerel_multitask_information_extraction

• Shared-transformer model for token-level NER and document/event multi-label classification.
• Documented metrics:
  • 746 documents
  • test token macro-F1 0.7318
  • test token macro-F1 all labels 0.7485
  • multi-label micro-F1 0.8035
  • validation threshold 0.56

Text-to-Image Product Search with Fine-Tuned CLIP

clip_product_search

• Vision-language retrieval workflow: CLIP fine-tuning, image embedding index and text-to-image catalog search.
• Documented metrics:
  • 39,742 train split
  • 4,416 test split
  • training subset 4,096
  • validation subset 768
  • baseline CLIP score 29.67
  • final validation CLIP score 30.51

Repository Structure

ai-research-projects/
  russian_llm_pretraining_and_sft/
  arxiv_semantic_retrieval/
  nerel_multitask_information_extraction/
  clip_product_search/

Reproducibility

Source datasets are not committed. Checkpoints, model weights, embedding indexes, generated outputs and local caches are also excluded from version control.

Each project README explains the reproduction assumptions for that experiment. Some projects require external datasets, GPU time, or notebook execution. Dataset licenses are controlled by the original dataset sources, not by this repository license.

Metrics are project-local experiment results and should not be treated as broad benchmarks.

Limitations

• These are small/local experiments rather than production systems.
• Reproduction depends on source data availability and compatible package versions.
• No model weights, checkpoints, embeddings or raw datasets are included.
• Some workflows are notebook-based.
• The reported metrics describe the documented project runs only; they do not imply general model or system quality.