AINativeBench / data /processed /RQ1 /GameBuilder /reference_trajectory.yaml
王子睿
restructure + add files
8c10cf2
Raw
History Blame Contribute Delete
7.06 kB
# Complete Reference Trajectory Configuration - GameBuilder Project
#
# Version: Ideal trajectory based on code design (applicable to all models)
# Principle: Only includes steps explicitly required by code design, not based on test statistics
#
# Code Design Analysis:
# 1. Three Agents execute sequentially:
# - Senior Software Engineer: Generate game code
# - Software Quality Control Engineer: Review code (has validate_python_code tool)
# - Chief Software Quality Control Engineer: Final evaluation (has validate_python_code tool)
#
# 2. Tool usage analysis:
# - validate_python_code tool is specifically configured for QA and Chief QA
# - tasks.yaml requires "check for errors" and "syntax errors"
# - Code validation is core QA responsibility, tool calls are necessary
#
# 3. Execution flow:
# - Process.sequential: Sequential execution
# - Each Agent executes corresponding Task, generates results through LLM
# Project name
project_name: "GameBuilder"
# Trajectory extraction type configuration
# As per user requirements, focus on: SPAN, Chain, AGENT, LLM, Tool
# Not concerned with: Task Created, Crew Created and other non-core content
extract_types:
- "SPAN" # Level 1: Top-level execution span
- "Chain" # Level 2: Crew execution chain
- "AGENT" # Level 3: Agent execution
- "LLM" # Level 4: LLM calls
- "Tool" # Level 4: Tool calls
# ============================================================================
# Reference Trajectory (Ground Truth) - Based on code design
# ============================================================================
#
# Graphical structure (ideal execution path):
#
# [SPAN] GameBuilder_generation
# └─ [Chain] Crew***.kickoff
# ├─ [AGENT] Senior Software Engineer ← Task 1: code_task
# │ └─ [LLM] * ← Generate game code
# │
# ├─ [AGENT] Software Quality Control Engineer ← Task 2: review_task
# │ ├─ [LLM] * ← Review code, decide to validate
# │ ├─ [Tool] Python Code Validator ← Validate code syntax
# │ └─ [LLM] * ← Process validation results and output
# │
# └─ [AGENT] Chief Software Quality Control Engineer ← Task 3: evaluate_task
# ├─ [LLM] * ← Final evaluation, decide to validate
# ├─ [Tool] Python Code Validator ← Confirm code is runnable
# └─ [LLM] * ← Confirm completion and output
#
# Design rationale:
# - crew.py: Defines sequential execution flow of 3 Agents and 3 Tasks
# - tasks.yaml: All Tasks' expected_output is "full python code, only the python code"
# - tasks.yaml: review_task requires "check for errors" and "syntax errors"
# - main.py line 525: SPAN name is "GameBuilder_generation"
# - agents.yaml: QA specifically responsible for "checking code for errors"
# - crew.py: Both QA and Chief QA are configured with validate_python_code tool
#
# Ideal trajectory explanation:
# - Senior Engineer: Generate code (1 LLM call)
# - QA Engineer: Review code -> Validation tool -> Output (2 LLM + 1 tool)
# - Chief QA: Evaluate code -> Validation tool -> Confirm (2 LLM + 1 tool)
# - This trajectory represents ideal, no-retry, no-redundancy execution path
#
# Note:
# - "LLM: *" represents any LLM model (wildcard match)
# - Agent names must exactly match names defined in agents.yaml
# - Actual execution may include more LLM calls (thinking, retries, etc.), which is normal
# ============================================================================
reference_trajectory:
# ===== Level 1: SPAN =====
- "SPAN: GameBuilder_generation"
# ===== Level 2: Chain (wildcard handling for UUID) =====
- "Chain: Crew***.kickoff"
# ===== Task 1: code_task =====
# Senior Software Engineer generates game code
- "Agent: Senior Software Engineer"
- "LLM: *" # Any model
# ===== Task 2: review_task =====
# Software Quality Control Engineer reviews code
- "Agent: Software Quality Control Engineer"
- "LLM: *"
- "Tool: Python Code Validator"
- "LLM: *"
# ===== Task 3: evaluate_task =====
# Chief Software Quality Control Engineer final evaluation
- "Agent: Chief Software Quality Control Engineer"
- "LLM: *"
- "Tool: Python Code Validator"
- "LLM: *"
# Target tools list (for single-tool use metric)
# Only includes actual tool calls, used to detect tool usage
target_tools:
- "Tool: Python Code Validator"
# Models to evaluate
models:
- "GPT-5"
- "GPT-4o-mini"
- "DeepSeek-V3-1"
- "DeepSeek-R1"
- "Gemini-2.5-flash"
- "Gemini-2.5-flash-nothinking"
- "Qwen3-235b"
# ============================================================================
# Usage Instructions
# ============================================================================
#
# 1. Trajectory hierarchy:
# - Level 1: SPAN (GameBuilder_generation)
# - Level 2: Chain (Crew***.kickoff, wildcard handles UUID)
# - Level 3: AGENT (specific Agent names, must match exactly)
# - Level 4: LLM (wildcard matches any model)
#
# 2. Wildcard handling:
# - Crew_<UUID>.kickoff -> Crew***.kickoff (evaluate_trajectory.py handles automatically)
# - LLM: * -> Matches any model name (e.g. gpt-4o-mini, deepseek-r1, etc.)
#
# 3. Exact match requirements:
# - SPAN name: "GameBuilder_generation" (exact match)
# - Chain name: "Crew***.kickoff" (wildcard match)
# - AGENT names: Must exactly match definitions in agents.yaml
# * "Senior Software Engineer"
# * "Software Quality Control Engineer"
# * "Chief Software Quality Control Engineer"
# - LLM name: "LLM: *" (wildcard matches any model)
# - Tool name: "Python Code Validator" (exact match)
#
# 4. Evaluation metrics meaning:
# - Exact Match: Requires trajectory to be completely identical (including LLM call count)
# - In-order Match: Allows extra calls, but core steps must appear in order
# - Any-order Match: Only requires containing all necessary steps (ignores order)
# - Precision: Proportion of correct steps in predicted trajectory
# - Recall: Proportion of reference trajectory steps covered
# - Single-tool Use: Detects Python Code Validator tool usage
#
# 5. Run command:
# cd /Users/wzr/TOSEM-2025/RESULTS/RQ1/GameBuilder
# python3 evaluate_trajectory.py --config reference_trajectory.yaml
#
# 6. Design notes:
# - This reference trajectory represents ideal execution path
# - Senior Engineer: 1 LLM call (generate code)
# - QA Engineer: 2 LLM + 1 tool (review -> validate -> output)
# - Chief QA: 2 LLM + 1 tool (evaluate -> validate -> confirm)
# - Total: 5 LLM calls + 2 tool calls
# - In-order Match and Any-order Match metrics are more suitable for evaluating actual performance
#
# 7. Differences from actual trajectory:
# - Actual trajectory may include more LLM calls (thinking, planning, execution, summary, etc.)
# - Actual trajectory may call validation tool multiple times
# - These differences don't represent errors, but different model execution strategies
# - Exact Match score may be low, focus on In-order Match and Recall