fable5-dataset 0.1.0

Load, preprocess, and manage the Fable5 agent trace datasets

Fable5 Dataset

Load, preprocess, and manage the Fable5 agent trace datasets for fine-tuning and evaluation.

Installation

pip install fable5-dataset

Dataset Sources

Source	Format	Description
Glint	Session-based with turns	Full agent sessions with tool use
armand0e	Conversation with tool_calls	Multi-turn conversations with function calling
vfable	Trajectory with tool_use	Agent trajectories with sequential tool use
Coding Excellence	Session-based with quality scores	High-quality coding sessions rated by experts
OpenCoven	Source/target pairs	Instruction-following input/output pairs
Victor	Prompt/response pairs	Single-turn coding instruction pairs

Quick Start

Load Datasets

# Load the Glint dataset
fable5 load glint

# Load all datasets with PII removal
fable5 load all --remove-pii

# Load with quality filter
fable5 load coding_excellence --min-quality 0.8 -o filtered.jsonl

View Statistics

# View stats for a specific dataset
fable5 stats --source glint

# View stats from a local file
fable5 stats traces.jsonl

# Compare all datasets
fable5 stats --source all

Convert Formats

# Convert to OpenAI chat format
fable5 convert traces.jsonl --format openai_chat -o train.jsonl

# Convert to Alpaca format
fable5 convert traces.jsonl --format alpaca -o alpaca.jsonl

Generate Benchmarks

# Generate 50 benchmark tasks from Glint
fable5 benchmark --source glint --num-tasks 50

# Generate category-specific benchmarks
fable5 benchmark --source coding_excellence --categories debugging implementation -o bench.jsonl

Split Data

# Split into 95/5 train/val
fable5 split traces.jsonl --train-ratio 0.95 --val-ratio 0.05

# Stratified split by tool distribution
fable5 split traces.jsonl --stratify-by tool --output-dir splits/

Programming API

from fable5_dataset import DatasetLoader, Preprocessor, BenchmarkGenerator, DatasetStats

# Load datasets
loader = DatasetLoader()
records = loader.load_dataset("glint", normalize=True, remove_pii=True)
all_data = loader.load_dataset("all")

# Preprocess
preprocessor = Preprocessor()
normalized = preprocessor.normalize_format(records, source_format="glint")
cleaned = preprocessor.remove_pii(normalized)
filtered = preprocessor.filter_quality(cleaned, min_quality=0.7)

# Statistics
stats = DatasetStats()
result = stats.compute_stats(records)
print(result.summary())
print(result.to_dict())

# Benchmark generation
gen = BenchmarkGenerator()
tasks = gen.generate_benchmark(records, num_tasks=50, categories=["debugging", "implementation"])
gen.save_benchmark(tasks, "benchmark.jsonl")

# Compare datasets
comparisons = stats.compare_datasets(all_data)
for name, ds_stats in comparisons.items():
    print(f"{name}: {ds_stats.total_rows} records, {ds_stats.avg_turns_per_session:.1f} avg turns")

License

MIT

Ecosystem

Part of the FableForge ecosystem — 21 open-source projects built from 210K real agent traces:

Project	Description
Anvil	Self-verified coding agent
VerifyLoop	Plan→Execute→Verify→Recover framework
ErrorRecovery	Self-healing middleware (3,725 error patterns)
FableForge-14B	The fine-tuned 14B model (4-stage training)
ShellWhisperer	1.5B edge agent (phone/RPi, 50ms)
ReasonCritic	Verification model (130 benchmark tasks)
TraceCompiler	Compile traces → LoRA skills
AgentRuntime	Persistent agent daemon (systemd for AI)
AgentSwarm	Multi-agent from real trace transitions
AgentTelemetry	Datadog for agents (token tracking, costs)
BenchAgent	HumanEval for tool-use (107 tasks)
AgentDev	VSCode extension with verification
TraceViz	Trace replay visualizer (Next.js)
AgentSkills	npm for agent behaviors
AgentCurriculum	5-stage progressive training
AgentFuzzer	Adversarial testing for agents
AgentConstitution	Safety guardrails from traces
CostOptimizer	Token cost reduction (50-80%)
AgentProfiler	Behavioral fingerprinting
TrajectoryDistiller	Trace→training data pipeline
Fable5-Dataset	HuggingFace dataset release