Stream bench pr

[sruan2]

Summary

• Adds a full benchmarking framework (eval/stream-bench/) for evaluating ACE on text-to-SQL tasks using the StreamBench suite, supporting the BIRD, CoSQL, and Spider datasets
• Extends ACE core (ace.py, reflector.py, llm.py, utils.py) with improvements needed for text-to-SQL evaluation: per-split data processors, SQL execution result passing to the reflector, more robust LLM error handling (HTTP 400 / invalid prompt recovery), and enriched run folder naming
• Adds data download, preprocessing, training, playbook evaluation, and plotting scripts with pre-built configuration files for common task sizes and difficulty balancing strategies

Changes

New: eval/stream-bench/

File	Purpose
run.py	Main ACE training/evaluation runner for text-to-SQL
run_playbook.py	Evaluate a saved playbook on any data split
data_processor.py	Data loading, schema formatting, and execution-based SQL accuracy
download_text2sql_data.py	Download raw BIRD/CoSQL/Spider SQLite databases
preprocess_streambench_{bird,cosql,spider}.py	Convert HuggingFace StreamBench data to .jsonl with schema
dataset_stats.py	Print dataset statistics
plot.py	Generate accuracy-over-time plots
analyze_logs.py	Analyze terminal output logs
data/bird_config.json	Pre-built BIRD task configurations
data/cosql_config.json	Pre-built CoSQL task configurations
data/spider_config.json	Pre-built Spider task configurations
README.md	Full setup and usage documentation

Modified: ACE core

• ace/ace.py: Support separate train_processor/val_processor/test_processor (backward-compatible with data_processor); pass metadata (db_name and curriculum) into run folder name; print correct/total counts alongside accuracy
• ace/core/reflector.py: Accept and format SQL execution results (predicted_result, ground_truth_result) and pass them into the reflector prompt for richer feedback
• ace/prompts/reflector.py: Add SQL execution results placeholder to both ground-truth and no-ground-truth reflector prompt templates
• llm.py: Detect and gracefully handle HTTP 400 / invalid-prompt / policy-violation errors without crashing; skip the affected sample and continue execution
• utils.py: Minor utilities supporting stream-bench evaluation