--- library_name: transformers license: mit base_model: openai-community/gpt2 language: - en tags: - modular-intelligence - text-generation - structured-reasoning - experimental --- # Modular Intelligence (GPT-2 baseline) This repository is an **experimental baseline** for **Modular Intelligence** built on top of `openai-community/gpt2`. The goal is **not** to claim that GPT-2 is “intelligent”, but to show how a **small, simple model** can be wrapped inside a **modular reasoning architecture**: - **Modules**: small, single-purpose “skills” (e.g. analysis note, strategy memo). - **Checkers**: strict reviewers that check the output of a module. - **Structured outputs**: fixed sections like CONTEXT / OPTIONS / RISKS / NEXT STEPS. Later, this same architecture can be reused with much stronger models. --- ## What this model is - A **GPT-2 checkpoint** configured as the engine behind a **Modular Intelligence** framework. - It is **not** heavily fine-tuned; it is used mainly to demonstrate: - Structured prompts - Module definitions - Checker patterns - Deterministic, repeatable formats Think of this repo as: > “The engine inside a modular reasoning system, using GPT-2 for a minimal, low-cost demo.” --- ## What’s different from base GPT-2? Base GPT-2 is a generic text generator. Here, GPT-2 is wrapped in a **specific contract**: 1. **Fixed module types** For example: - `analysis_note_v1` - `document_explainer_v1` - `strategy_memo_v1` - `message_reply_v1` - `profile_application_v1` - `system_blueprint_v1` - `modular_brainstorm_v1` 2. **Fixed output sections** Each module must respond in a strict, labelled format. Example (Strategy Memo): - CONTEXT - OBJECTIVE - CONSTRAINTS - OPTIONS - RECOMMENDATION - RISKS - NEXT_ACTIONS 3. **Paired checkers** Certain modules have a checker module that: - Re-reads the original task - Reviews the draft output - Returns a verdict + issues + suggested fixes 4. **Use pattern** Instead of “just generating text”, you: - Call a **module** with structured inputs - Get a **structured output** - Optionally call a **checker** on that output So the “intelligence” here is in the **architecture and prompts**, not in new weights. --- ## Dataset This repository **does not introduce** a new training dataset and **does not re-train** GPT-2. - **Base model**: `openai-community/gpt2` - **Training objective**: next-token prediction (causal language modeling) - **Original GPT-2 pretraining data** (by OpenAI, not included here): - Large, general-domain English web corpus (“WebText”) - ~40 GB of text from web pages linked from Reddit posts with score ≥ 3 - Mixed content (news, blogs, forums, technical/non-technical) In this repository: - GPT-2 is used **as-is** as the language engine. - The **Modular Intelligence** behaviour comes from: - The **module prompts** (how we talk to the model) - The **checker prompts** (how we review its answers) - The **fixed output formats** No new datasets are uploaded or used for further fine-tuning inside this repo. --- ## Modular Intelligence: modules and checkers (simple view) ### Generator modules Each generator is a “skill” with a fixed format. 1. **Analysis Note (`analysis_note_v1`)** - **Inputs**: - `context` – short description of the situation or text - `questions` – what you want to understand - `constraints` – any limits (time, style, scope) - **Outputs (sections)**: - CONTEXT - QUESTIONS - FRAMEWORK - ANALYSIS - CONCLUSION - NEXT_STEPS 2. **Document Explainer (`document_explainer_v1`)** - **Inputs**: - `document_text` - `focus` - `audience` - **Outputs**: - SNAPSHOT - KEY_POINTS - STRUCTURE - DETAILED_EXPLANATION - IMPLICATIONS - ACTIONS 3. **Strategy Memo (`strategy_memo_v1`)** - **Inputs**: - `context` - `objective` - `constraints` - **Outputs**: - CONTEXT - OBJECTIVE - CONSTRAINTS - OPTIONS - RECOMMENDATION - RISKS - NEXT_ACTIONS 4. **Message / Post Reply (`message_reply_v1`)** - **Inputs**: - `source_text` - `your_angle` - `tone_notes` - **Outputs**: - DRAFT_REPLY 5. **Profile / Application Draft (`profile_application_v1`)** - **Inputs**: - `target_role_or_goal` - `your_background` - `audience` - **Outputs**: - POSITIONING - KEY_POINTS - FULL_DRAFT 6. **System / Architecture Blueprint (`system_blueprint_v1`)** - **Inputs**: - `objective` - `current_state` - `constraints` - **Outputs**: - OBJECTIVE - CURRENT_STATE - COMPONENTS - FLOWS - RISKS - IMPROVEMENTS - NEXT_STEPS 7. **Modular Brainstorm (`modular_brainstorm_v1`)** - **Inputs**: - `problem_or_domain` - `goal` - **Outputs**: - OBJECTIVE - CURRENT - MODULES - CHECKERS - DATA_NEEDS - NEXT_STEPS --- ### Checker modules Checkers are “reviewers” that inspect generated outputs. Examples: 1. **Analysis Note Checker (`analysis_note_checker_v1`)** - **Inputs**: - `original_task` - `draft_output` - **Outputs**: - VERDICT - STRUCTURE - CLARITY - ALIGNMENT - GAPS - FIXES 2. **Document Explainer Checker (`document_explainer_checker_v1`)** - VERDICT - ACCURACY - STRUCTURE - AUDIENCE_FIT - MISSING - FIXES 3. **Strategy Memo Checker (`strategy_memo_checker_v1`)** - VERDICT - OBJECTIVE_ALIGNMENT - CONSTRAINT_HANDLING - OPTION_QUALITY - RISKS - FIXES 4. **Style & Voice Checker (`style_voice_checker_v1`)** - VERDICT - STYLE_MATCH - TONE - REDUNDANCY - SUGGESTIONS 5. **Profile Checker (`profile_checker_v1`)** - VERDICT - ALIGNMENT - SIGNAL - CLARITY - FIXES 6. **System Checker (`system_blueprint_checker_v1`)** - VERDICT - COHERENCE - GAPS - FLOW_ISSUES - RISKS - FIXES --- ## How to use this model (simple) You can treat this model like any GPT-2 text generator, **but** if you want Modular Intelligence behaviour: 1. Pick a module (e.g. `strategy_memo_v1`). 2. Build a prompt that: - States the module name - Lists the inputs clearly - Lists the required output sections 3. Ask the model to **fill in each section in order**. 4. Optionally call the corresponding checker with: - Original task - Draft output A reference implementation and UI are provided in the accompanying Hugging Face Space (if linked), but the pattern can be re-implemented in any environment. --- ## Limitations - GPT-2 is **small and outdated** by modern standards. - It will: - Hallucinate - Get facts wrong - Sometimes ignore structure - Struggle with long contexts The goal is to demonstrate the **architecture**, not to claim state-of-the-art performance. Do **not** use this model for high-stakes decisions or any application where mistakes could cause real harm. --- ## License and IP - Code and configuration: **MIT License**. - The **Modular Intelligence architecture, module definitions, and checker patterns** are a conceptual layer that can be reused and extended, but the name and approach may be treated as separate intellectual property by the author. Always review the base model’s license (`openai-community/gpt2`) for any additional constraints.