tensorboard-ai 0.2.0

TensorBoard plugin with conversational AI capabilities

TensorBoard AI

A TensorBoard plugin that adds conversational AI analysis to your training dashboards. Analyze metrics, compare runs, and generate custom visualizations — all through natural language.

Quick Start

git clone https://github.com/jackyoung96/tensorboard-ai.git
cd tensorboard-ai
pip install .

Set your API key:

export ANTHROPIC_API_KEY="your-key"     # Anthropic (default)
# or
export OPENAI_API_KEY="your-key"        # OpenAI
# or use AWS IAM credentials             # Bedrock

Launch TensorBoard:

tensorboard --logdir /path/to/logs

Open the AI tab in TensorBoard. Click Analyze All or use the AI button on any chart.

Features

AI-Powered Training Analysis

• One-click Analyze All for comprehensive analysis of all visible metrics
• Per-metric AI button on each chart card for focused analysis
• Automatic training type detection (SFT, RL/PPO, DPO, Pre-training) with domain-specific insights
• Cross-metric correlation analysis (e.g., grad_norm spike when reward drops)
• Hyperparameter-aware analysis — automatically reads hparams and highlights differences across runs

Multi-Provider LLM Support

• Anthropic Claude (default) — including extended thinking mode
• OpenAI (GPT-4o, GPT-5 series)
• AWS Bedrock — automatic region prefix for cross-region model IDs
• Runtime provider/model switching via sidebar UI (no restart needed)

Interactive Chat Panel

• Always-visible sidebar with SSE streaming responses
• Markdown rendering with GFM support (tables, code blocks, lists)
• Clickable metric names in responses — scrolls to and highlights the corresponding chart
• Follow-up conversations with full context retention
• Analysis history: switch between previous analyses

Chart Visualization

• Plotly-based interactive charts with TensorBoard-like styling
• EMA smoothing with adjustable weight slider
• Horizontal axis modes: Step / Relative time / Wall time
• AI-generated custom Plotly visualizations via conversation

Token Optimization

• Smart data sampling: uniform grid + anomaly detection + curvature-based points (max 50 per series)
• Constant value detection — outputs constant=X instead of full data
• Run name alias compression (common prefix/suffix removal, UUID/hash stripping)
• Duplicate metric deduplication: identical data across tags grouped together

Dashboard Layout

Three-panel layout with resizable panels:

Settings Sidebar	Chart Grid	AI Analysis Panel
Chart settings, analysis settings, run selection, tag filter	Interactive Plotly charts with per-metric AI buttons	Streaming chat with markdown, metric links, history

Settings

• Chart: Smoothing slider, horizontal axis mode (Step / Relative / Wall)
• Analysis: Language (Auto/EN/KO/JA/ZH/ES/FR/DE), provider, model, extended thinking, debug mode
• Runs: Checkbox list with color swatches, regex filter, Select All/None
• Tag Filter: Filter visible charts by tag name

Configuration

Environment Variables

Variable	Description	Default
TENSORBOARD_AI_PROVIDER	LLM provider (anthropic, openai, bedrock)	anthropic
TENSORBOARD_AI_MODEL	Model name override	Provider default
ANTHROPIC_API_KEY	Anthropic API key	—
OPENAI_API_KEY	OpenAI API key	—
AWS_ACCESS_KEY_ID	AWS access key (Bedrock)	IAM role
AWS_SECRET_ACCESS_KEY	AWS secret key (Bedrock)	IAM role
AWS_SESSION_TOKEN	AWS session token (Bedrock)	—
AWS_REGION	AWS region (Bedrock)	us-east-1

CLI Options

# Standard usage
tensorboard --logdir /path/to/logs

# Enable debug mode (shows raw prompt data in AI panel)
tensorboard --logdir /path/to/logs --tensorboard_ai_debug

Development

# Install in development mode
pip install -e ".[dev]"

# Run tests
pytest

# Lint & format
ruff check .
ruff format .

# Type checking
mypy tensorboard_ai/

# Build package
python -m build

Architecture

tensorboard_ai/
├── plugin.py              # TBPlugin + TBLoader — routes, CLI flags
├── backend/
│   ├── server.py          # SSE streaming chat endpoint
│   ├── data_access.py     # Reads scalars, hyperparameters via DataProvider
│   └── chart_gen.py       # Validates LLM-generated Plotly specs
├── providers/
│   ├── base.py            # Abstract provider protocol + registry
│   ├── anthropic.py       # Claude integration
│   ├── openai.py          # OpenAI integration
│   └── bedrock.py         # AWS Bedrock integration
├── prompts/
│   └── system.py          # System prompts with training type detection
└── static/
    ├── index.js           # Dashboard layout, settings, chart grid
    ├── chat.js            # Chat UI, SSE client, markdown rendering
    └── chart_renderer.js  # Plotly charts, data sampling, run aliases

License

Apache-2.0