How to Build a Stable Pipeline on the Fable 5 Traces Dataset in Google Colab

The Fable 5 Traces dataset appeared on Hugging Face and has already attracted the attention of researchers: it stores real traces of agent sessions from Claude Fable 5 — an Anthropic model designed for tool use. MarkTechPost published a detailed tutorial on how to build a stable pipeline around it in Google Colab.

What is Fable 5 Traces

Fable 5 is one of the latest models in the Claude family, specialized in agent tasks. The traces dataset captures how the model thinks when invoking functions: tool requests, their arguments, returned results, and chains-of-thought. Each record in the JSONL file represents a complete session: from system prompt to final answer. The structure is nested and heterogeneous — different traces can contain different fields, which creates the main technical complexity when working with the dataset.

• System prompt and user request
• List of tool calls with arguments and results
• Model CoT-reasoning (if enabled)
• Final answer to the user

How the Pipeline Works in Colab

The tutorial authors deliberately avoided high-level libraries — in particular, Hugging Face's `datasets` — in favor of manual parsing using Python's standard `json` module. The logic is simple: Colab regularly updates library versions, and code that depends on specific versions of `datasets` or `transformers` breaks easily without warning. Manual JSONL parsing is resilient to such surprises.

The pipeline goes through several stages. First, inspection: reviewing the repository file structure, checking the number of records, identifying anomalies. Then normalization — bringing tool calls from different traces to a unified schema so they can be processed programmatically. A separate step is data auditing for secrets. Agent session traces often contain API keys, tokens, or other sensitive data that leaked into the context through tools. The tutorial shows how to automatically detect and redact such fragments before using the dataset for training.

After cleanup, the pipeline creates visualizations: distribution of tool call types, session lengths, frequency of specific functions. These plots help quickly assess data quality and balance.

Basic Classifier Instead of Neural Network

The final part of the tutorial covers export and training. The authors prepare two versions of the dataset: the complete version (with CoT-reasoning) and the truncated no-CoT version (requests and responses only, without chains of thought). The second option is safer for fine-tuning: CoT-reasoning may contain internal artifacts unsuitable for training external models.

A naive Bayes classifier is trained on the prepared data, written in pure Python — without PyTorch, TensorFlow, or other heavy frameworks. Such a baseline solves several tasks:

• Quickly check if the data contains statistically significant patterns
• Provide a baseline for comparing neural network models
• Run experiments without GPU and long wait times
• Identify class imbalance at the hypothesis verification stage

"We parse the merged JSONL file manually to keep

Colab reliable and predictable" — from the MarkTechPost tutorial.

What This Means

Public datasets with real agent traces are rare. Most training sets for LLMs contain synthetic data or web text, but not records of real agent reasoning during tool invocation. Fable 5 Traces fills this gap and opens opportunities for studying agent behavior in Claude models.

The tutorial lowers the barrier to entry: all code runs in a browser in a few minutes without complex environment setup. For developers who want to fine-tune their own models on agent tasks or study tool use patterns, it's a ready-made starting point.