App Pages & Features
Dashboard
- High-level metrics, quick links into Generate/Judge/Uploads/Vector DB.
- Recent activity and status badges for running pipelines.
Generate
- Configure provider/model, sampling params (temperature/top-p/top-k/repetition penalty/max tokens), and concurrency.
- Honors dataset profiles for schema/labels; outputs flow to Dataset and can be judged/embedded.
Profiles
- Select built-in profiles or custom ones; shows schema fields and evaluation axes.
- Template Builder UI for adding/editing profiles (see Data Profiles doc for JSON structure).
Reverse Engineer
- Upload single/batch binaries; get static/dynamic analysis and JSONL export.
- Progress per file; optional sandbox report upload; see safety notes in Reverse Engineering doc.
Transcribe
- YouTube URL or local audio → in-browser Whisper.js for transcription/translation/summarization.
- Supports progress, streaming decode, and optional embedding into vector DB.
Deduplication
- Exact hash and near-dup (shingle + Jaccard) detection with adjustable cutoff and neighbor caps.
- Cluster view with sample previews; activity log for dedupe runs.
Stratified Mix
- Set target distributions across length, domain, and difficulty; normalize targets to 100%.
- Computes current dataset strata and produces a guidance hint (
stratifiedNextAsk) for underrepresented bins.
- Can auto-inject the hint into Teacher LLM calls to steer future generations toward gaps.
Uploads
- Drag/drop files or folders (PDF/DOCX/TXT/MD/HTML/JSON).
- Chunk/overlap controls, embedding model selection, optional Qdrant sync, per-file progress/errors.
Embeddings
- Inspect embeddings, sample from Qdrant, visualize scatter/heatmap, and build context from selections.
- Sync status indicators and pagination for large sets.
Dataset
- Browse/edit generated and ingested samples with schema awareness from the active profile.
- Inline editing of instruction/input/output/metadata; filters/search by profile.
- Grammar editor with AST viewer and test input playground.
- Toggle grammar usage for generation; validate grammars before use.
Environment
- Set API keys/endpoints (OpenAI, Gemini, OpenRouter, vLLM host, Ollama, custom OpenAI-compatible).
- Concurrency knobs, chunking mode, embedding model selection, and danger-zone data reset.
LLM as a Judge
- Configure judge/teacher models and concurrency.
- Run evaluations across accuracy/completeness/clarity/overall (or profile-specific axes); results feed Close-the-loop.
Close the Loop
- Automates regenerate → evaluate cycles to improve samples based on judge scores.
- Monitors runs with status and aggregated metrics.
Vector DB Integration
- Configure Qdrant host/API key/collection; ensure collection exists with correct vector size.
- Trigger resample/load-more from Qdrant; view sync health.
Fine-Tune Export
- Export curated datasets to training-ready formats (e.g., JSONL) using the active profile schema.
- Include metadata fields and evaluation scores if desired.