DOCSEARCHbyAlgolia
Submit Docs

NanoResearch

repository·main·Indexed 23 days ago

https://github.com/openraiser/nanoresearch

An autonomous AI-driven research engine (v0.1.0) that automates the end-to-end scientific process from ideation and literature review to experimental execution and LaTeX paper drafting. Unlike standard writing tools, it executes actual computational experiments on local GPUs or SLURM clusters, generates real data-backed figures, and supports a self-evolving 'Evo' pipeline. Features include a Python CLI with TUI mode, Codex integration, and a Feishu (Lark) robot for remote pipeline management.

Tokens
242.7K
Snippets
488
Records
884
Agent score
81%

What's inside nanoresearch

  1. Overview of the Autoresearch skill

    main

    The autoresearch skill orchestrates end-to-end autonomous AI research projects using a two-loop architecture. It manages the full lifecycle from literature survey to paper writing by maintaining structured state, running experiment-synthesis cycles, and routing to domain-specific skills. It is designed to run fully autonomously, making progress without constant user intervention while providing periodic progress reports (HTML/PDF).

    Core Workflow:

    1. Bootstrap: Scope question, search literature, and form initial hypotheses.
    2. Inner Loop: Rapid, autonomous experiment iterations (Pick hypothesis $\rightarrow$ experiment $\rightarrow$ measure $\rightarrow$ record $\rightarrow$ learn).
    3. Outer Loop: Periodic reflection to synthesize results, find patterns, and update the research direction.
    4. Finalize: Write the final paper and presentation.

    Use this skill when starting a new research project, running autonomous experiments, or managing multi-hypothesis research efforts.

  2. What is NanoResearch

    main

    NanoResearch is an end-to-end autonomous AI research engine designed to automate the entire scientific research process—from initial research ideas to a complete, data-backed LaTeX paper.

    Unlike traditional AI writing tools, NanoResearch executes actual computational experiments. It generates runnable code, submits it to local GPUs or SLURM clusters, collects real experimental results, and uses that data to generate figures and write the paper. This ensures that every table, chart, and data point in the final output is derived from real training logs rather than LLM hallucinations.

  3. Overview of Unsloth features and capabilities

    main

    Unsloth is an open-source framework designed for efficient LLM fine-tuning and reinforcement learning (RL). It supports training, running, evaluating, and saving models like Llama, DeepSeek, Qwen, Mistral, and Gemma.

    Key Capabilities:

    • Efficient Training: Supports full-finetuning, pretraining, and 4-bit, 8-bit, or 16-bit training. It is optimized to be up to 2x faster with 70% less VRAM.
    • Reinforcement Learning: Provides an efficient RL library (using 80% less VRAM) supporting algorithms like GRPO and GSPO.
    • Broad Model Support: Supports vision (VLM), text-to-speech (TTS), BERT, and multimodal models. Any model compatible with transformers can be used in Unsloth.
    • Hardware Support: Works on Linux, Windows, Colab, Kaggle, and supports NVIDIA, AMD, and Intel GPUs.
    • Deployment: Facilitates exporting and integration with inference engines such as Ollama, llama.cpp, and vLLM.

  4. Key features of NanoResearch

    main

    NanoResearch provides several advanced capabilities for autonomous research:

    • Real Experiments: Auto-submits GPU/SLURM jobs and collects real metrics.
    • Multi-Model Routing: Configure different LLMs for different pipeline stages (e.g., different models for coding vs. writing).
    • Checkpoint & Resume: Recover from failures at any stage.
    • Evidence-Grounded Writing: Ensures paper content is tied to actual experiment results.
    • Auto Figure Gen: Generates visualizations from genuine experiment artifacts.
    • Multi-Format Templates: One-click switching between NeurIPS, ICML, and arXiv LaTeX templates.
    • Multi-Interface Control: Accessible via CLI (TUI or streaming logs), Claude Code, or Feishu Bot.

  5. Use the Unsloth Skill for fast fine-tuning

    main

    The Unsloth skill provides expert guidance for high-performance LLM fine-tuning. Use this skill when you need to implement 2-5x faster training, reduce memory usage by 50-80%, or optimize LoRA/QLoRA workflows for models like Llama, Mistral, Gemma, and Qwen.

    Trigger this skill when:

    • Working with the unsloth library.
    • Asking about Unsloth features or APIs.
    • Implementing fine-tuning solutions.
    • Debugging Unsloth-specific code.
    • Learning best practices for memory-efficient training.

  6. NanoResearch use cases

    main

    Developers and researchers can use NanoResearch for:

    • Research Prototyping: Quickly turning an idea into a full experiment and paper workspace.
    • Autonomous Experiments: Letting the system handle code generation, GPU training, and result analysis.
    • Benchmark Generation: Batch-running multiple topics with reproducible results.
    • Paper Draft Assistance: Producing LaTeX drafts grounded in real experimental data.
    • Research Audit Trails: Using workspaces, manifests, and stage artifacts as an auditable research log.

  7. Fine-tune LLMs with Unsloth

    main

    Unsloth is an open-source framework designed for LLM fine-tuning and reinforcement learning (RL).

    Core Capabilities:

    • Fine-tuning: Learn basics, best practices, and how to prepare datasets.
    • Hyperparameter Optimization: Guidance on LoRA rank, alpha, epochs, batch size, gradient accumulation, and QLoRA vs LoRA.
    • Reinforcement Learning (RL): Support for GRPO (to train reasoning models like DeepSeek-R1), DPO, ORPO, KTO, and GSPO.
    • Vision/Multimodal Fine-tuning: Training Vision-Language Models (VLM) via GRPO and RL.
    • Continued Pretraining: Also known as Continued Finetuning, allowing models to learn new languages.
    • Quantization-Aware Training (QAT): Quantize models to 4-bit with Unsloth and PyTorch to recover accuracy.

  8. Use the brainstorming-research-ideas skill for research ideation

    main

    The brainstorming-research-ideas skill provides structured ideation frameworks to help researchers move from vague curiosity to concrete, defensible research proposals. It offers ten complementary ideation lenses targeting different cognitive modes.

    When to use this skill:

    • Starting a new research direction and needing structured exploration.
    • Feeling stuck on a current project and seeking fresh angles.
    • Evaluating the potential of a half-formed idea.
    • Preparing for brainstorming sessions with collaborators.
    • Transitioning between research areas to find high-leverage entry points.
    • Reviewing a field to identify underexplored gaps.

    When NOT to use this skill:

    • If you already have a well-defined research question and need execution guidance.
    • If you need help with experimental design or methodology (use domain-specific skills instead).
    • If you need a literature review (use scientific-skills:literature-review instead).

  9. What is Unsloth Dynamic 2.0 quantization?

    main

    Unsloth Dynamic 2.0 is an advanced quantization method for GGUFs that outperforms standard methods in benchmarks like 5-shot MMLU and KL Divergence.

    Key features include:

    • Intelligent Layer Selection: Dynamically adjusts the quantization type for every possible layer in a model, rather than just modifying select layers.
    • Broad Compatibility: Works on all model architectures, including both MoE (Mixture of Experts) and non-MoE models.
    • Custom-tailored Schemes: Each model (e.g., Gemma 3 vs. Llama 4) uses a unique quantization scheme optimized for its specific architecture.
    • Expanded Formats: Includes Q4_NL, Q5.1, Q5.0, Q4.1, and Q4.0 formats to maximize efficiency on Apple Silicon and ARM devices.
    • High-Quality Calibration: Uses a calibration dataset of >1.5M high-quality, hand-curated tokens to enhance conversational chat performance.

  10. What is Unsloth Dynamic Quantization?

    main

    Unsloth Dynamic Quantization is a method of selectively quantizing layers of an LLM to achieve extreme compression without catastrophic loss in quality.

    Instead of quantizing all layers uniformly, it makes important layers higher precision (e.g., 8 or 16 bits) and less important layers lower precision (e.g., 1, 2, 3, 4, 5, or 6 bits). This approach is particularly effective for Mixture-of-Experts (MoE) models and can be paired with an imatrix calibration dataset designed for chat and coding performance.

  11. Identify LLM training data formats

    main

    LLM training datasets require specific formats depending on the training objective. Choose a format based on your goal:

    FormatDescriptionTraining Type
    Raw CorpusRaw text (websites, books, articles)Continued Pretraining (CPT)
    InstructInstructions and target output examplesSupervised fine-tuning (SFT)
    ConversationMulti-turn dialogues between user and AISupervised fine-tuning (SFT)
    RLHFConversations with ranked responsesReinforcement Learning (RL)

  12. Reinforcement Learning (RL) capabilities in Unsloth

    main

    Unsloth provides tools to transform standard models into reasoning models using Reinforcement Learning (RL).

    Key Capabilities:

    • Model Scale: With 15GB VRAM, you can transform models up to 17B parameters (e.g., Llama 3.1 8B, Phi-4 14B, Mistral 7B, or Qwen2.5 7B).
    • Vision/Multimodal Support: Unsloth supports RL for Vision/multimodal models.
    • Low VRAM Requirements: A minimum of 5GB VRAM is sufficient to train reasoning models for any model with 1.5B parameters or less.

    Important Note on RL Success: RL relies on receiving signals from both good and bad answers to 'prune' the model's output distribution. If the probability of a correct answer is always 0, RL will not work. It is recommended to perform RL starting from an already instruction-finetuned model to ensure the probability of following instructions is above 0.