Danamam - TensorTrainer

Danamam - TensorTrainer is a browser-only trainer for a small conversational buddy. You add question–answer pairs, train a lightweight classifier on top of sentence embeddings, then chat and optionally export the trained weights as JSON.

No server-side AI is required: TensorFlow.js and Universal Sentence Encoder run entirely in the client.

---

Features

• Q/A training pairs with add, list, and remove (before training)
• Semantic encoding of questions (512-dim embeddings)
• Small classifier (dense + dropout + softmax) trained in the browser
• Chat after training, with optional session history and clear
• Export trained model as danamam-model.json (architecture, weights, answer list, metadata)
• Nuxt 3, Vue 3 (Composition API), Nuxt UI, toasts, and loading states

---

How it works

1. On load, the app downloads Universal Sentence Encoder (first visit may take a while).
2. You enter Question and Answer fields and click Add pair. Repeat; you need at least two pairs to enable Start training.
3. Start training embeds all questions, one-hot encodes answers (one class per distinct answer string), and fits a small tf.sequential model for 100 epochs with a validation split.
4. After training, use Chat with your buddy; the app embeds your message and picks the softmax argmax over your answer set.
5. Download model (JSON) saves everything needed to reconstruct the buddy offline (you would still need compatible loader code and the USE model for inference).

This is answer selection, not open-ended text generation: replies always come from answers you trained on.

---

Tech stack

Layer	Technology
Framework	Nuxt 3
UI	Nuxt UI + Tailwind CSS
ML	TensorFlow.js, @tensorflow-models/universal-sentence-encoder

---

Prerequisites

• Node.js 18+ (20 LTS recommended)
• A modern desktop browser (training and embedding are CPU/WebGPU dependent; large batches can feel slow on low-end devices)

---

Installation

gh repo clone alihd-tech/TensorTrainer
cd /{dir}
pnpm install

---

Development

pnpm run dev

Open http://localhost:3000. The main UI is defined in app.vue (root app component).

---

Build and preview

pnpm run build
pnpm run preview

---

Export format (danamam-model.json)

The download includes:

• architecture — model.toJSON() from TensorFlow.js
• weights — serialized weight tensors
• answers — ordered list of unique answer strings (output index alignment)
• metadata — for example ISO timestamp and pair count

Importing this file back into an app is not implemented in this repository; the export is for backup, experimentation, or a custom loader you build.

---

Project layout

app.vue              # Main trainer + chat UI and TF.js logic
nuxt.config.ts       # Nuxt config (SSR noExternal for TF packages)
assets/css/main.css  # Tailwind entry
public/logo.svg      # Brand logo
tailwind.config.js   # Theme extensions (primary / secondary)
package.json

There is no pages/index.vue in this project; routing defaults to the root app.vue experience.

---

Limitations

• Replies are limited to answers seen in training (classification over your answer set).
• Minimum two Q/A pairs.
• Distinct answers define the number of classes; duplicate answer text across questions shares one class.
• Large datasets or weak hardware can make training and first encoder load slow or memory-heavy.

---

Possible extensions

• Confidence threshold and an “I don’t know” bucket
• Persist pairs in localStorage
• Import previously exported JSON for inference-only mode
• Dataset editor (CSV import, reorder, bulk delete)

Screenshot

---

License

MIT License