Convergent-7B-data

# Convergent-7B Training Data <p align="center"> <img src="convergent-banner.jpg" alt="Convergent-7B — bigcompute.science research companion" width="800"> </p> **Training data for the [bigcompute.science](https://bigcompute.science) research companion model.** > **Early Preview** — This dataset is a work in progress. It is expressly designed to train a research assistant for the [bigcompute.science](https://bigcompute.science) MCP server as part of the Convergent conjecture-driven GPU research project. The dataset will be updated frequently as new experiments, findings, and tool definitions are added. Expect changes to schema, tool names, and content until we reach a GA release. The complete training dataset used to fine-tune [cahlen/Convergent-7B](https://huggingface.co/cahlen/Convergent-7B). | Repository | Description | |------------|-------------| | **[cahlen/Convergent-7B](https://huggingface.co/cahlen/Convergent-7B)** | Trained model weights | | **[cahlen/Convergent-7B-data](https://huggingface.co/datasets/cahlen/Convergent-7B-data)** | This repo — training dataset | | **[cahlen/convergent](https://github.com/cahlen/convergent)** | Training code, eval, CLI toolkit | ## Dataset Description 5,799 training entries in ChatML message format (cleaned and deduplicated), covering: - **Computational number theory**: continued fractions, Zaremba's conjecture, Hausdorff dimensions, Kronecker coefficients, Ramsey numbers, Flint Hills series, Cohen-Lenstra heuristics - **Agentic tool calling**: Hermes-format function calls to the bigcompute.science MCP server, including multi-turn ReAct trajectories - **CUDA kernel development**: GPU programming for number theory with architecture-specific optimization - **Research methodology**: proof strategies, experiment design, student guidance - **Synthetic reasoning**: Deep mathematical Chain-of-Thought from Qwen2.5-Math-72B and creative synthesis from Gemma-4-26B ## Format Each entry is a JSON object with a `messages` array in ChatML format: ```json { "messages": [ {"role": "system", "content": "You are Convergent, the bigcompute.science research companion..."}, {"role": "user", "content": "How many Zaremba exceptions exist for digit set {1,2,3}?"}, {"role": "assistant", "content": "<tool_call>\n{\"name\": \"get_zaremba_exceptions\", \"arguments\": {}}\n</tool_call>"} ] } ``` Multi-turn entries include `tool` role messages for agentic ReAct trajectories: ```json { "messages": [ {"role": "system", "content": "..."}, {"role": "user", "content": "Check the Zaremba verification status"}, {"role": "assistant", "content": "<tool_call>...</tool_call>"}, {"role": "tool", "content": "{\"status\": \"completed\", \"exceptions\": 0}"}, {"role": "assistant", "content": "The verification is complete with zero exceptions..."} ] } ``` ## Composition | Source | Entries | Description | |--------|---------|-------------| | Curated domain blocks (40+ modules) | ~1,150 | Identity, tool calling (23 MCP tools), nvcc-validated CUDA, number theory, error recovery, paper comprehension, student guidance | | Qwen2.5-Math-72B (synthetic) | ~3,100 | Deep mathematical reasoning and Chain-of-Thought | | Gemma-4-26B (synthetic) | ~1,200 | Creative synthesis, experiment design, long-form reasoning | | Hermes FC (external) | 300 | Diverse tool-calling patterns from NousResearch | | **Total (after dedup + cleaning)** | **5,799** | Off-topic entries and near-duplicates removed | ### Category Breakdown | Category | Count | Percentage | |----------|-------|------------| | Mathematical reasoning (CoT) | ~3,500 | 60% | | Tool-calling (agentic) | ~710 | 12% | | Knowledge / factual | ~800 | 14% | | Multi-turn conversations | ~520 | 9% | | CUDA code generation | ~270 | 5% | ## Data Sources See [DATA_SOURCES.md](https://github.com/cahlen/convergent/blob/main/DATA_SOURCES.md) for complete documentation of all sources, including: - Cahlen Humphreys' paper on prime convergents of continued fractions - Boise State University and Florida Atlantic University number theory research - Open Erdős problems - NVIDIA GPU architecture specifications - NousResearch/hermes-function-calling-v1 - bigcompute.science experimental findings ## Generation Pipeline The training toolkit is open-source: [github.com/cahlen/convergent](https://github.com/cahlen/convergent) ```bash ./convergent generate-blocks # Generate curated domain training blocks ./convergent generate-synthetic # Generate synthetic data from remote LLMs ./convergent merge # Merge, deduplicate, remove eval leaks ./convergent validate # Validate format and quality ``` ## License CC-BY-4.0 — You are free to share and adapt this dataset with attribution. ## Links - [bigcompute.science](https://bigcompute.science) — Conjecture-driven GPU research in computational mathematics - [MCP Server](https://mcp.bigcompute.science) — Model Context Protocol server for experimental data and tools - [Convergent-7B Model](https://huggingface.co/cahlen/Convergent-7B) — Trained model weights on HuggingFace - [Training Toolkit](https://github.com/cahlen/convergent) — Full pipeline source code on GitHub - [guerrillamathematics.com](https://guerrillamathematics.com) — Mathematical research blog ## Citation ```bibtex @misc{humphreys2026convergent, author = {Humphreys, Cahlen}, title = {Convergent-7B Training Data: Computational Number Theory for Agentic Research}, year = {2026}, url = {https://huggingface.co/datasets/cahlen/Convergent-7B-data} } ``` --- *This project is maintained by a single person. If you run into issues, please file them on [GitHub](https://github.com/cahlen/convergent/issues) or [HuggingFace](https://huggingface.co/cahlen/Convergent-7B/discussions) and I will do my best to address them. I apologize in advance for any delays in response time.*