| # 🧠 ThoughtSwitch V1 1.7B Instruct — A Mode-Adaptive Reasoning Language Model | |
| > **Model ID**: [`BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct`](https://huggingface.co/BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct) | |
| > **Architecture**: Decoder-only transformer (GPT-style) | |
| > **Parameters**: 1.7 Billion | |
| > **Capabilities**: Dynamic "Thinking" vs. "Non-Thinking" mode-switching | |
| > **Fine-Tuned for**: Instruction-following | |
| --- | |
| ## 🚀 Overview | |
| **ThoughtSwitch V1** is a next-generation instruction-tuned language model that brings a new paradigm to text generation: **Autonomous Cognitive Mode Switching**. | |
| It is capable of **interpreting user prompts and switching between two distinct modes** of behavior: | |
| - 🧠 **Thinking Mode**: Deep reasoning, logical step-by-step solutions, slow but deliberate outputs. | |
| - 💬 **Non-Thinking Mode**: Quick completions, casual replies, storytelling, and chat-like fluency. | |
| Whether you're building reasoning agents, fast assistants, or multi-modal chains-of-thought applications, ThoughtSwitch adapts intelligently—so **you don’t have to force the prompt**. | |
| --- | |
| ## 🧠 Key Features | |
| - ✅ **Autonomous Mode Switching** | |
| Understands when to think deeply and when to generate fluently, based on prompt phrasing. | |
| - ✅ **Instruction Tuned** | |
| Trained to follow human-like instructions and align closely with user intent. | |
| - ✅ **1.7B Parameters** | |
| Small enough for efficient inference, yet powerful for sophisticated reasoning. | |
| - ✅ **Open Weights** | |
| Fully accessible under a permissive license (specify in HF model card). | |
| --- | |
| ## ✨ Example Prompts | |
| Prompt (Thinking Mode): | |
| "Think step by step to solve this math problem: What is 17 multiplied by 23?" | |
| → Reasoned output with intermediate steps and justification. | |
| Prompt (Non-Thinking Mode): | |
| "Write a quick sci-fi story about a robot discovering love." | |
| → Smooth, creative storytelling without unnecessary reasoning. | |
| --- | |
| ## 🔧 Usage | |
| from transformers import AutoTokenizer, AutoModelForCausalLM | |
| tokenizer = AutoTokenizer.from_pretrained("BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct") | |
| model = AutoModelForCausalLM.from_pretrained("BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct") | |
| prompt = "Think step by step: Why does ice float on water?" | |
| inputs = tokenizer(prompt, return_tensors="pt") | |
| outputs = model.generate(**inputs, max_new_tokens=150) | |
| print(tokenizer.decode(outputs[0], skip_special_tokens=True)) | |
| --- | |
| ## 🧪 Intended Use Cases | |
| - 🧠 **Reasoning Agents** — For multi-hop question answering, logical puzzles, or decision support. | |
| - 📚 **Tutoring & Education** — Adaptive explanations that vary depth based on student prompts. | |
| - 🤖 **Conversational AI** — More natural and flexible interactions with variable "thinking effort". | |
| - ✍️ **Creative Writing** — Generate stories, poems, and ideas with or without deep context. | |
| --- | |
| ## ⚠️ Limitations | |
| - Like all LLMs, it may hallucinate or generate biased content. | |
| - Mode switching is **probabilistic**, not guaranteed—prompt clearly for best results. | |
| - Performance may vary outside of English or unfamiliar domains. | |
| --- | |
| ## 📈 Performance (Unofficial Benchmarks) | |
| | Task | Performance | | |
| |------------------------|------------------| | |
| | Commonsense Reasoning | ✅ Strong | | |
| | Instruction Following | ✅ Strong | | |
| | Fast Casual Generation | ✅ Very Strong | | |
| | Math (Step-by-Step) | ⚠️ Moderate | | |
| | Factual QA | ⚠️ May hallucinate| | |
| --- | |
| ## 🛠️ Model Details | |
| - **Architecture**: GPT-style decoder (causal LM) | |
| - **Training**: Custom pretraining with hybrid reasoning/non-reasoning dataset | |
| - **Instruction Fine-Tuning**: Yes, using curated prompt-response pairs | |
| - **Token Limit**: 2048 tokens (extendable with rope scaling) | |
| --- | |
| ## 🔍 Quantized Version | |
| Looking for fast inference? | |
| Check out the GGUF-quantized version (by @mradermacher) for compatibility with llama.cpp, KoboldAI, and other lightweight runtimes. | |
| --- | |
| ## 📄 Citation | |
| If you use this model in your research or application, please cite it as: | |
| @misc{thoughswitch2025, | |
| title={ThoughtSwitch V1 1.7B Instruct: A Mode-Adaptive Reasoning Language Model}, | |
| author={BrainWave-ML}, | |
| year={2025}, | |
| howpublished={\url{https://huggingface.co/BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct}} | |
| } | |
| --- | |
| ## 💬 Contact | |
| For issues, feedback, or collaboration: | |
| - 🤖 Hugging Face Page: https://huggingface.co/BrainWave-ML/ThoughtSwitch-V1-1.7b-Instruct | |
| - 📧 Email: *[YourContact@domain.com]* | |
| - 🌐 Website: *[https://brainwave-ml.ai]* (optional) | |
| - 💬 Discord or Community: *Coming Soon* | |
| --- | |
| ## 🙌 Acknowledgments | |
| Developed by the team at **BrainWave-ML**. Inspired by the question: | |
| *“What if language models could choose when to think?”* | |
| --- | |
| > **ThoughtSwitch**: Think when you need to. Generate when you don't. | |