A research-first firm building custom data, models and evaluation
for problems an off-the-shelf API cannot solve.
π About us
Mercity Research Labs works with product companies and enterprises on custom training, new model architectures, optimization and private research. We handle every stage, from the training data to the evaluation, and our clients own the weights and the code.
We are research-first and work in the open. The datasets, weights and tools on this page come straight out of our own projects, published so anyone can use them and reproduce our results.
π§ͺ What we focus on
Synthetic data generation. We have generated hundreds of billions of tokens of synthetic data and trained models on it. Every dataset ships with its generation code, filters and a coverage report.
Custom architecture and training. When an LLM or API model isn't the right fit, we design and train models around the task, including modified attention, custom kernels, pruning and distillation (for example, Qwen3-8B pruned from 36 layers to 30).
Evaluation and benchmarking. Every project starts with a domain-specific benchmark, and every change is measured against it. We read the outputs ourselves.
π¦ Highlights on the Hub
| π‘οΈ Topical Guardrails | SAE-based classifiers on Gemma 1B / 4B for keeping models on topic |
| βΈοΈ Figure Skating Classification | Five pose-based architectures (CNN+BiLSTM, Transformer, CTM, Transformer+BiLSTM, GCN) trained on SkatingVerse |
| βοΈ creative-writing-llm | Text-generation model tuned for creative writing |
| π§ ReasonBridge-URT | ~1M-row reasoning dataset |
| π FineWeb-Preprocessed | 10M-document preprocessed pretraining corpus |
| πΌοΈ laion-subset | Curated image-text subset of LAION |
Browse everything: models Β· datasets Β· collections
π οΈ Open-source tools
| Tool | Stack | What it does |
|---|---|---|
| Simula | Python | Deliberate synthetic data generation: describe the distribution as a spec, and Simula plans, generates and verifies a corpus against it |
| PromptKeep | Python | Versioned prompt templates with lineage, plus a drop-in OpenAI wrapper that records every run |
π€ Work with us
Have a problem that doesn't fit the API? If an off-the-shelf model gets you most of the way and the rest is what your business depends on, we should talk.