Skip to main content

​
Overview

The prime eval tui command launches an interactive terminal user interface for browsing evaluation results. It automatically discovers and displays all evaluation runs from your workspace.

​
Usage

prime eval tui [OPTIONS]

​
Options

No required options. The TUI will auto-discover results from standard locations.

​
What It Shows

The TUI provides a hierarchical browser:
  1. Environment selection - All environments with completed evaluations
  2. Model selection - All models evaluated for that environment
  3. Run selection - All evaluation runs for that environment + model combo
  4. Rollout viewer - Individual prompts, completions, and metrics

​
Discovery

Results are discovered from:
  • ./outputs/evals/ - Global output directory
  • ./environments/*/outputs/evals/ - Per-environment output directories
Each run must have both:
  • results.jsonl - Rollout data
  • metadata.json - Evaluation metadata

​
Environment Selection Screen

┌─ Select Environment ─────────────────────────────────┐
│ gsm8k - Models: 2, Runs: 5                          │
│ math-python - Models: 1, Runs: 3                    │
│ alphabet-sort - Models: 2, Runs: 4                  │
└──────────────────────────────────────────────────────┘

q: Quit  Enter: Select
Keys:
  • ↑/↓ or j/k - Navigate
  • Enter - Select environment
  • q - Quit

​
Model Selection Screen

┌─ Environment: gsm8k ─────────────────────────────────┐
│ Select Model                                         │
│ openai/gpt-4.1-mini - Runs: 3                       │
│ anthropic/claude-sonnet-4 - Runs: 2                 │
└──────────────────────────────────────────────────────┘

q: Quit  b: Back  Enter: Select
Keys:
  • ↑/↓ - Navigate
  • Enter - Select model
  • b or Backspace - Go back
  • q - Quit

​
Run Selection Screen

┌─ Environment: gsm8k ─────────────────────────────────┐
│ Model: openai/gpt-4.1-mini                          │
│ Select Run                                           │
│ abc123 - 2026-03-03 14:23 | Reward: 0.867           │
│ def456 - 2026-03-02 10:15 | Reward: 0.823           │
│ ghi789 - 2026-03-01 16:47 | Reward: 0.891           │
└──────────────────────────────────────────────────────┘

┌─ Run Details ────────────────────────────────────────┐
│ Run ID: abc123                                       │
│ Environment: gsm8k                                   │
│ Model: openai/gpt-4.1-mini                          │
│ Avg reward: 0.867                                    │
│ Runtime: 2m 34.5s                                    │
└──────────────────────────────────────────────────────┘

q: Quit  b: Back  Enter: Select
Keys:
  • ↑/↓ - Navigate runs
  • Enter - View rollout details
  • b or Backspace - Go back
  • q - Quit

​
Rollout Viewer

The main screen shows prompts, completions, and metrics side-by-side: 
┌─ Metadata ───────────────────────────────────────────┐
│ Environment: gsm8k      Record: 1/30                 │
│ Model: openai/gpt-4.1-mini                          │
│ Run ID: abc123          Examples: 10                 │
│ Date: 2026-03-03 14:23  Rollouts/ex: 3              │
└──────────────────────────────────────────────────────┘

┌─ Prompt ─────────────┬─ Completion ──────────────────┐
│ user: What is 2+2?   │ assistant: The answer is 4.   │
│                      │                                │
│                      │ tool call: calculate          │
│                      │ {"expression": "2+2"}         │
│                      │                                │
│                      │ tool result: 4                │
│                      │                                │
│                      │ assistant: The result is 4.   │
└──────────────────────┴────────────────────────────────┘

┌─ Details ────────────────────────────────────────────┐
│ Reward: 1.000                                        │
│ Answer: 4                                            │
│ Info: {"calculation": "2+2=4"}                      │
└──────────────────────────────────────────────────────┘

q: Quit  b: Back  ←/→: Prev/Next  s: Search  c: Copy
Keys:
  • ←/→ or h/l - Navigate between rollouts
  • s - Search prompt/completion text
  • c - Enter copy mode
  • b or Backspace - Go back to run list
  • q - Quit
  • d - Toggle dark/light theme

​
Search Mode

Press s to search within prompts and completions: 
┌─ Search (regex, case-insensitive) ───────────────────┐
│ [calculate               ]                           │
└──────────────────────────────────────────────────────┘

┌─ Prompt results (0) ────┬─ Completion results (2) ───┐
│                          │   245 | tool call: calculate│
│                          │   312 | assistant: The calcu│
└──────────────────────────┴────────────────────────────┘

Esc: Close  Enter: Select  ←/→: Switch column
Keys:
  • Type to search (regex supported)
  • ↑/↓ - Navigate results
  • ←/→ - Switch between prompt and completion results
  • Enter - Jump to selected match
  • Esc - Close search
Matches are highlighted for 3 seconds after selection.

​
Copy Mode

Press c to enter copy mode: 
┌─ Copy Mode ──────────────────────────────────────────┐
│ Tab: switch columns                                  │
│ Highlight text with mouse drag or Shift+Arrow       │
│ Esc: close                                           │
└──────────────────────────────────────────────────────┘

┌─ Prompt ─────────────┬─ Completion ──────────────────┐
│ user: What is 2+2?   │ assistant: The answer is 4.   │
│                      │                                │
│ [Text is selectable] │ [Text is selectable]          │
└──────────────────────┴────────────────────────────────┘

q: Quit  Tab: Next column  c: Copy  Esc: Close
Keys:
  • Tab - Switch between prompt and completion
  • Mouse drag or Shift+Arrow - Select text
  • c - Copy selected text to clipboard
  • Esc - Exit copy mode
  • q - Quit

​
Display Features

​
Message Formatting

Messages are formatted with role-based styling:
  • user messages - Standard text
  • assistant messages - assistant: prefix in bold
  • tool calls - tool call: prefix with function name and arguments
  • tool results - tool result: prefix in dimmed style
  • errors - Red text with error: prefix

​
Metrics Display

The details panel shows:
  • Reward - Scalar reward from rubric (formatted to 3 decimals)
  • Answer - Ground truth answer from task
  • Info - Additional environment-specific data (formatted as JSON)
  • Task - Full task data if available

​
Lazy Loading

Results are loaded lazily for performance:
  • File handles opened on-demand
  • Lines read as needed
  • Metadata count used when available
  • Caching for already-read records
This allows the TUI to handle evaluations with thousands of rollouts efficiently.

​
Themes

Toggle between dark and light themes with d:
  • black-warm (default) - Dark theme with warm accent colors
  • white-warm - Light theme with matching warm tones

​
Examples

​
Basic Usage

# Run an evaluation
prime eval run gsm8k -m gpt-4.1-mini -n 10 -s

# Launch TUI
prime eval tui

​
View Specific Results

The TUI automatically finds all results, so just launch it: 
prime eval tui
Navigate to your environment → model → run.

​
Search for Patterns

  1. Launch TUI and navigate to a run
  2. Press s to open search
  3. Type a regex pattern (e.g., error|failed)
  4. Navigate results with arrow keys
  5. Press Enter to jump to a match

​
Copy Completions

  1. Navigate to a rollout
  2. Press c to enter copy mode
  3. Tab to completion column
  4. Select text with mouse or Shift+Arrow
  5. Press c to copy to clipboard

​
File Locations

Results are saved by prime eval run --save-results to: 
./outputs/evals/
└── gsm8k--openai--gpt-4.1-mini/
    └── abc123/
        ├── results.jsonl
        └── metadata.json
Or per-environment: 
./environments/gsm8k/outputs/evals/
└── gsm8k--openai--gpt-4.1-mini/
    └── abc123/
        ├── results.jsonl
        └── metadata.json
The TUI scans both locations.

​
Performance

The TUI is optimized for large evaluations:
  • Lazy file reading - Only loads visible data
  • Incremental parsing - Reads JSONL line-by-line
  • Metadata caching - Avoids re-parsing metadata files
  • Efficient rendering - Textual’s virtual DOM
Evaluations with 10,000+ rollouts are handled smoothly.

​
Troubleshooting

​
No Evaluations Found

┌─ Select Environment ─────────────────────────────────┐
│ No completed evals found                            │
└──────────────────────────────────────────────────────┘
Solution: Run an evaluation with --save-results: 
prime eval run gsm8k -m gpt-4.1-mini -n 5 -s
prime eval tui

​
Corrupted Results

If results.jsonl is malformed, that rollout will show as {}. Solution: Check the file manually: 
cat outputs/evals/gsm8k--openai--gpt-4.1-mini/abc123/results.jsonl | jq

​
Terminal Size

If the TUI appears cramped, resize your terminal: 
resize -s 40 160  # 40 rows, 160 columns
prime eval tui
Recommended minimum: 24 rows × 80 columns

​
Search Not Working

Search uses regex with case-insensitive matching. Test your pattern: 
echo "test string" | grep -iE 'pattern'
If the pattern is invalid, an error appears below the search box.

​
Keyboard Reference

​
Global

  • q - Quit application
  • d - Toggle dark/light theme
  • ↑/↓ or j/k - Move selection
  • Enter - Select item
  • b or Backspace - Go back one screen

​
Rollout Viewer

  • ←/→ or h/l - Previous/next rollout
  • s - Open search
  • c - Enter copy mode
  • b or Backspace - Return to run list

​
Search Mode

  • Type - Enter search pattern
  • ↑/↓ - Navigate results
  • ←/→ - Switch prompt/completion
  • Enter - Jump to selected match
  • Esc - Close search

​
Copy Mode

  • Tab or Shift+Tab - Switch column
  • Mouse drag - Select text
  • Shift+Arrow - Select text (keyboard)
  • c - Copy to clipboard
  • Esc - Exit copy mode

​
Tips

  • Use search (s) to quickly find errors or specific patterns
  • Copy mode (c) allows extracting full completions for analysis
  • Results persist across runs - view historical evaluations anytime
  • The TUI works great with tmux/screen for remote evaluation monitoring
  • Use --state-columns when running evals to save additional fields visible in the TUI

Build docs developers (and LLMs) love

Get started for freeTalk to us