lzy510016411/fable5-gpt5.5-opus4.7-mixed-agent-traces

Fable5 · GPT-5.5 · Opus-4.7 Mixed Agent Traces

A high-density post-training mixture for agentic reasoning, instruction following, code generation, function calling, and tool-use decision making.

This is the training-data release behind [Qwen3.5-9B-Distill-Agent-Instruct](https://huggingface.co/lzy510016411/qwen3.5-9b-distill-agent-instruct), an Agent Instruct model distilled and post-trained from Qwen3.5-9B-Base. The title highlights three of the mixture's principal model-labelled trajectory families—Claude Fable5, GPT-5.5 Agent, and Claude Opus 4.7—while mixed reflects the additional GLM-5.2 Agent, Qwen3.7-Max, Glaive, Hermes, When2Call, and code-instruction components. Rather than being a raw conversation dump, the dataset is a deliberately composed mixture of long-horizon agent trajectories, step-level tool-use examples, function-calling conversations, code instructions, trace-inversion samples, and balanced tool/no-tool decisions.

The released files contain 20,409 training records, 211,580 message events, and 67,865 explicit tool-call steps. Of all records, 11,580 include tool definitions. Statistics are calculated directly from this release.

What is included

The mixture combines the following trajectory families (names follow the released file identifiers):

Together, these components teach more than tool syntax. They target the complete agent loop: understanding an instruction, deciding whether a tool is necessary, selecting and parameterizing a tool, consuming observations, continuing across multiple steps, recovering from intermediate state, and producing a useful final answer.

Author-curated cleaning pipeline

Every subset was passed through an author-curated, task-aware cleaning and mixture-construction pipeline. Depending on the source, the pipeline includes:

• —Conversation-schema normalization across system, user, assistant, tool-call, and tool-response events.
• —Trajectory integrity checks to remove malformed records and preserve executable call/response ordering.
• —Step-level slicing at meaningful decision boundaries, including pre-tool-response and final-assistant cut points.
• —Judge-gated filtering for step datasets, retaining accepted samples together with judge reasons and evidence.
• —Source-aware sampling to prevent a small number of origins from dominating the mixture.
• —Structural and content deduplication for repeated prompts, traces, and near-identical agent paths.
• —Tool-use rebalancing so the model learns both when to invoke a function and when a direct answer is the better action.
• —Mixture-level balancing across agentic reasoning, coding, function calling, and general instruction following.

This dedicated cleaning process is the main distinction of the release: noisy trajectories are not treated as equally valuable tokens. The mixture emphasizes valid decision points, coherent state transitions, diverse tool behavior, and high-information supervision.

Data format

Most files use JSON Lines with a messages array and an optional tools field:

{
  "tools": ["... optional tool definitions ..."],
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."},
    {"role": "tool_call", "content": "..."},
    {"role": "tool_response", "content": "..."}
  ]
}

Step-level files additionally expose provenance and filtering metadata such as source_id, step_index, target_roles, cut_reason, label, judge_reason, judge_model, and judge_evidence. The individual components are intentionally retained as separate files so users can reproduce the mixture, ablate a trajectory family, or build their own sampling policy.

Intended use

This dataset is designed for supervised post-training of language models, especially for:

• —instruction-following and conversational alignment;
• —agentic, multi-step tool use;
• —function calling and structured action generation;
• —tool-use boundary learning (“call or answer”);
• —code generation and task completion;
• —data-mixture and trajectory-ablation research.

It is training data, not an evaluation benchmark. Users should create source-separated validation sets and normalize role/tool schemas for their target training framework.

Notes and limitations

• —The release is heterogeneous by design; schemas and available metadata differ across components.
• —Model/provider names in file identifiers are provenance labels supplied with the trajectory groups and do not imply endorsement by those providers.
• —Automated filtering substantially improves consistency but does not guarantee that every sample is correct, safe, or suitable for every deployment.
• —Users are responsible for reviewing the applicable terms and licenses of upstream data components before redistribution or commercial use.

Curator

Collected, cleaned, deduplicated, filtered, rebalanced, and assembled by [lzy510016411](https://huggingface.co/lzy510016411) for the post-training of Qwen3.5-9B-Distill-Agent-Instruct.