Configuration - SkyDiscover

Configuration File Structure

SkyDiscover uses YAML files to configure all aspects of the discovery process:

config.yaml

# General settings
max_iterations: 100
checkpoint_interval: 10
log_level: "INFO"

# LLM configuration
llm:
  models:
    - name: "gpt-5"
      weight: 1.0
  temperature: 0.7
  max_tokens: 32000

# Search algorithm
search:
  type: "adaevolve"
  num_context_programs: 4
  database:
    population_size: 20
    num_islands: 2

# Prompt configuration
prompt:
  system_message: "You are an expert to help find the best solution."

# Evaluator configuration
evaluator:
  timeout: 360
  cascade_evaluation: true

General Settings

max_iterations

int

default:"100"

Maximum number of discovery iterations

checkpoint_interval

int

default:"10"

Save checkpoint every N iterations

log_level

string

default:"INFO"

Logging level: DEBUG, INFO, WARNING, ERROR, CRITICAL

log_dir

string

default:"null"

Custom directory for log files (default: output_dir/logs)

language

string

default:"python"

Target language: python for code, text for prompts

file_suffix

string

default:".py"

File extension for generated programs (.py, .txt, .cpp, etc.)

diff_based_generation

bool

default:"true"

Generate diffs instead of full programs (faster, more focused)

max_solution_length

int

default:"60000"

Maximum characters in generated solutions

max_parallel_iterations

int

default:"1"

Number of iterations to run concurrently (1 = sequential)

LLM Configuration

Basic Model Setup

llm:
  models:
    - name: "gpt-5"
      weight: 1.0
  
  api_base: "https://api.openai.com/v1"
  temperature: 0.7
  top_p: 0.95
  max_tokens: 32000
  timeout: 600
  retries: 3
  retry_delay: 5

Multiple Models

SkyDiscover samples from multiple models based on weights:

llm:
  models:
    - name: "gpt-5"
      weight: 0.5
    - name: "gemini/gemini-3-pro"
      weight: 0.3
    - name: "claude-3-7-sonnet"
      weight: 0.2
  
  temperature: 0.7
  max_tokens: 32000

Per-Model Configuration

llm:
  models:
    - name: "gpt-5"
      weight: 0.7
      temperature: 0.8
      max_tokens: 40000
    - name: "gemini/gemini-3-pro"
      weight: 0.3
      temperature: 0.6
      max_tokens: 30000
      api_key: ${GEMINI_API_KEY}  # Model-specific key
  
  # Defaults for models without explicit values
  temperature: 0.7
  max_tokens: 32000

Evaluator and Guide Models

llm:
  # Main models for solution generation
  models:
    - name: "gpt-5"
      weight: 1.0
  
  # Models for LLM-as-judge evaluation
  evaluator_models:
    - name: "gpt-4o-mini"
      weight: 1.0
  
  # Models for high-level strategy (paradigm breakthrough)
  guide_models:
    - name: "o1"
      reasoning_effort: "high"
      weight: 1.0

If evaluator_models or guide_models are not specified, they default to the same models as the main models list.

LLM Parameters

llm.api_base

string

default:"https://api.openai.com/v1"

Base URL for API requests

llm.api_key

string

default:"null"

API key (defaults to environment variable)

llm.temperature

float

default:"0.7"

Sampling temperature (0.0 = deterministic, 2.0 = very random)

llm.top_p

float

default:"0.95"

Nucleus sampling threshold

llm.max_tokens

int

default:"32000"

Maximum tokens per generation

llm.timeout

int

default:"600"

Request timeout in seconds

llm.retries

int

default:"3"

Number of retry attempts on failure

llm.retry_delay

int

default:"5"

Seconds to wait between retries

llm.reasoning_effort

string

default:"null"

Reasoning effort for o1/o3 models: low, medium, high

Search Configuration

Top-K (Simple)

search:
  type: "topk"
  num_context_programs: 4
  database:
    db_path: "database.db"
    log_prompts: true

AdaEvolve (Adaptive Multi-Island)

search:
  type: "adaevolve"
  num_context_programs: 4
  database:
    # Population
    population_size: 20
    num_islands: 2
    
    # Adaptive intensity
    decay: 0.9
    intensity_min: 0.15
    intensity_max: 0.5
    use_adaptive_search: true
    
    # Island selection
    use_ucb_selection: true
    
    # Migration
    use_migration: true
    migration_interval: 15
    migration_count: 5
    
    # Archive
    use_unified_archive: true
    fitness_weight: 1.0
    novelty_weight: 0.0
    diversity_strategy: "code"  # "code", "metric", or "hybrid"
    
    # Dynamic islands
    use_dynamic_islands: true
    max_islands: 5
    spawn_productivity_threshold: 0.015
    spawn_cooldown_iterations: 30
    
    # Paradigm breakthrough
    use_paradigm_breakthrough: true
    paradigm_window_size: 10
    paradigm_improvement_threshold: 0.12
    paradigm_max_uses: 2
    paradigm_num_to_generate: 3

AdaEvolve Parameters Explained

Adaptive Intensity:

decay: EMA weight (0.9 = slow adaptation)
intensity_min: Min search intensity (exploitation)
intensity_max: Max search intensity (exploration)

Migration:

Every migration_interval iterations, top migration_count programs migrate to neighboring islands

Diversity Strategy:

code: Edit distance between solutions
metric: Euclidean distance in metric space
hybrid: Weighted combination

Paradigm Breakthrough:

Triggered when improvement rate drops below threshold
Generates high-level strategy ideas to escape local optima

Beam Search

search:
  type: "beam_search"
  num_context_programs: 4
  database:
    beam_width: 5
    beam_selection_strategy: "diversity_weighted"
    beam_diversity_weight: 0.3
    beam_temperature: 1.0
    beam_depth_penalty: 0.0

Best-of-N

search:
  type: "best_of_n"
  num_context_programs: 4
  database:
    best_of_n: 5

EvoX (Co-Evolution)

search:
  type: "evox"
  num_context_programs: 4
  database:
    database_file_path: null  # Uses default search strategy
    evaluation_file: null     # Uses default evaluator
    config_path: null         # Uses default config
    auto_generate_variation_operators: true

OpenEvolve Native (MAP-Elites)

search:
  type: "openevolve_native"
  num_context_programs: 4
  database:
    num_islands: 5
    population_size: 40
    archive_size: 100
    exploration_ratio: 0.2
    exploitation_ratio: 0.7
    elite_selection_ratio: 0.1
    feature_dimensions: ["complexity", "diversity"]
    feature_bins: 10
    migration_interval: 10
    migration_rate: 0.1

GEPA Native

search:
  type: "gepa_native"
  num_context_programs: 4
  database:
    population_size: 40
    candidate_selection_strategy: "epsilon_greedy"  # "best", "pareto"
    epsilon: 0.1
    acceptance_gating: true
    use_merge: true
    merge_after_stagnation: 15

Prompt Configuration

System Message

prompt:
  system_message: |
    You are an expert mathematician specializing in circle packing.
    
    Your task is to improve a function that arranges N circles in a unit
    square to maximize the sum of their radii.
    
    Key insights:
    - Hexagonal patterns achieve densest packing
    - Edge effects make square containers harder
    - Consider layered/shell arrangements

External File

prompt:
  system_message: "system_prompt.txt"  # Load from file

Template Selection

prompt:
  template: "default"      # "default" or "evox"
  template_dir: null       # Custom template directory
  system_message: "system_message"
  evaluator_system_message: "evaluator_system_message"  # For LLM judge

Simplification Suggestion

prompt:
  suggest_simplification_after_chars: 500  # Suggest simplifying if solution > 500 chars

Evaluator Configuration

evaluator:
  timeout: 360              # Evaluation timeout (seconds)
  max_retries: 3            # Retry on failure
  
  # Cascade evaluation
  cascade_evaluation: true
  cascade_thresholds: [0.3, 0.6]
  
  # LLM-as-a-judge
  llm_as_judge: false

evaluator.timeout

int

default:"360"

Maximum seconds for evaluation

evaluator.max_retries

int

default:"3"

Retry failed evaluations N times

evaluator.cascade_evaluation

bool

default:"true"

Enable multi-stage evaluation (requires evaluate_stage1() and evaluate_stage2() in evaluator)

evaluator.cascade_thresholds

array

default:"[0.3, 0.6]"

Score thresholds for cascade stages

evaluator.llm_as_judge

bool

default:"false"

Enable LLM judge for qualitative feedback

Agentic Configuration

Enable codebase-aware solution generation:

agentic:
  enabled: false
  codebase_root: null  # Auto-detected from initial_program location
  
  # Agent loop limits
  max_steps: 5
  per_step_timeout: 60.0
  overall_timeout: 300.0
  
  # Context management
  max_context_chars: 400000
  max_file_chars: 50000
  max_search_results: 50
  max_files_read: 20
  
  # Regex safety
  regex_timeout: 2.0
  max_regex_length: 200
  
  # Repo map
  repo_map_max_depth: 4
  
  # File access
  allowed_extensions:
    - ".py"
    - ".txt"
    - ".md"
    - ".json"
    - ".yaml"
  excluded_dirs:
    - ".git"
    - "__pycache__"
    - "node_modules"
    - ".venv"

Monitor Configuration

Enable live dashboard:

monitor:
  enabled: true
  host: "127.0.0.1"
  port: 8765
  max_solution_length: 10000
  
  # AI summary settings
  summary_model: "gpt-5-mini"
  summary_api_key: null  # Defaults to OPENAI_API_KEY
  summary_api_base: "https://api.openai.com/v1"
  summary_top_k: 3       # Summarize top-K programs
  summary_interval: 0    # Auto-generate every N programs (0 = manual only)

See the Monitoring guide for details.

Human Feedback Configuration

human_feedback_enabled: true
human_feedback_file: "human_feedback.md"
human_feedback_mode: "append"  # "append" or "replace"

Write feedback in Markdown:

human_feedback.md

# Iteration 15

The packing is too dense in the center. Try spreading circles more evenly.

# Iteration 22

Good improvement! Now focus on optimizing corner utilization.

Environment Variable Expansion

Use ${VAR} syntax to reference environment variables:

llm:
  api_key: ${OPENAI_API_KEY}
  models:
    - name: "gemini/gemini-3-pro"
      api_key: ${GEMINI_API_KEY}

Example Configurations

max_iterations: 100
checkpoint_interval: 10
log_level: "INFO"

llm:
  models:
    - name: "gpt-5"
      weight: 1.0
  temperature: 0.7
  max_tokens: 32000

search:
  type: "topk"
  num_context_programs: 4

prompt:
  system_message: "You are an expert to help find the best solution."

evaluator:
  timeout: 360
  cascade_evaluation: false

diff_based_generation: true

Next Steps

Model Providers

Set up different LLM providers

Writing Evaluators

Create effective scoring functions

Monitoring

Enable the live dashboard

Benchmarks

See real configuration examples

​Configuration File Structure

​General Settings

​LLM Configuration

​Basic Model Setup

​Multiple Models

​Per-Model Configuration

​Evaluator and Guide Models

​LLM Parameters

​Search Configuration

​Top-K (Simple)

​AdaEvolve (Adaptive Multi-Island)

​Beam Search

​Best-of-N

​EvoX (Co-Evolution)

​OpenEvolve Native (MAP-Elites)

​GEPA Native

​Prompt Configuration

​System Message

​External File

​Template Selection

​Simplification Suggestion

​Evaluator Configuration

​Agentic Configuration

​Monitor Configuration

​Human Feedback Configuration

​Environment Variable Expansion

​Example Configurations

​Next Steps

Model Providers

Writing Evaluators

Monitoring

Benchmarks

Configuration File Structure

General Settings

LLM Configuration

Basic Model Setup

Multiple Models

Per-Model Configuration

Evaluator and Guide Models

LLM Parameters

Search Configuration

Top-K (Simple)

AdaEvolve (Adaptive Multi-Island)

Beam Search

Best-of-N

EvoX (Co-Evolution)

OpenEvolve Native (MAP-Elites)

GEPA Native

Prompt Configuration

System Message

External File

Template Selection

Simplification Suggestion

Evaluator Configuration

Agentic Configuration

Monitor Configuration

Human Feedback Configuration

Environment Variable Expansion

Example Configurations

Next Steps