README.md

---
pretty_name: Claude Fable 5 Agent Traces
license: cc-by-4.0
language:
  - en
annotations_creators:
  - machine-generated
task_categories:
  - text-generation
size_categories:
  - 10K<n<100K
tags:
  - "traces"
  - "code"
  - "agentic"
  - "tool-use"
  - "coding-agent"
  - "coding-agents"
  - "agent-traces"
  - "instruction-following"
  - "function-calling"
  - "parallel-tool-calling"
  - "multi-turn"
  - "sft"
  - "distillation"
  - "reasoning"
  - "chain-of-thought"
  - "cot"
---

# Claude Fable 5 Agent Traces

![Moonshiner — Claude Fable 5 instruction following, tool use, and coding](./moonshiner-dataset-banner.png)

<div align="center">
  <h2>2,161 TRAJECTORIES · 11,235 TRAINING ROWS · 11 MB PARQUET · 656 MB JSONL</h2>
</div>

> Generated by **[moonshiner](https://github.com/greghavens/moonshiner)** — an open harness for
> distilling verified instruction-following, tool-use, and agentic coding traces.


Behavior-preserving **instruction-following, tool-use, and agent trajectories**
from **Claude Fable 5** (`anthropic/claude-fable-5`). The category and row-share tables
below describe the actual mix seen during training rather than assuming a
particular task domain.

**This is an actively growing dataset. More is coming**: additional training
programs and substantially more sessions will be added to this same repo.

## What makes it different

- **All real model trajectories.** Every row is a cumulative prefix of a genuine
  Claude Fable 5 session captured end-to-end. The agent's causal exploration, tool
  arguments, results, corrections, and final responses are retained.
- **One next step per row.** A trajectory with N assistant turns produces N
  rows. Row k contains the complete context through assistant turn k; that final
  assistant message is the sole training target.
- **Runtime-normalized.** Runtime plumbing, UI decoration, control sequences,
  and verbose success boilerplate are removed or canonicalized while causal
  context remains.
- **Independently verified.** Coding sessions must pass deterministic tests and
  protected-file checks. Instruction-following sessions must pass deterministic
  tool-call, staging, argument, and response-constraint checks. Every retained
  trajectory also clears independent review.

- **Reasoning-effort step-down.** Failed trace attempts proceed through `xhigh → medium → low` (up to the configured attempt count) and stop at the first judge-accepted trace. If higher reasoning fails a task that lower reasoning succeeds on, the lower-effort trace is retained.

## Task mix

High-level training programs, calculated from accepted trajectories using the
same program mapping published in the seed catalog:

| kind | trajectories | share | row share | flavor |
|---|---:|---:|---:|---|
| Tool calling | 640 | 29.6% | 12.4% | Select tools, construct grounded arguments, run independent calls together, and stage dependent calls. |
| Instruction following | 545 | 25.2% | 12.4% | Honor constraints, formats, corrections, state, context, memory, relevance, and abstention. |
| Building | 286 | 13.2% | 14.0% | Implement complete libraries, services, CLIs, workflows, and systems from specifications. |
| Debugging | 230 | 10.6% | 10.5% | Diagnose failures, repair defects, resolve compiler/runtime issues, and preserve regressions. |
| Project & integration | 138 | 6.4% | 11.6% | Coordinate multi-file, repository-scale, migration, and integration work. |
| Clarification | 101 | 4.7% | 1.7% | Recognize missing information, ask only when required, and never invent parameters. |
| Feature development | 81 | 3.7% | 5.3% | Extend working systems while preserving existing behavior. |
| Error recovery | 70 | 3.2% | 1.9% | Recover from tool failures, partial results, retries, and idempotency hazards. |
| Seed authoring | 49 | 2.3% | 29.1% | Accepted trajectories grouped by their explicit published dataset category. |
| Refactoring & performance | 21 | 1.0% | 1.0% | Restructure safely and improve measured performance without behavior drift. |

## Languages (current drop)

`English`, `Python`, `TypeScript`, `Go`, `PowerShell`, `Java`, `Rust`, `C#`, `Ruby`, `Bash`, `Zsh`, `JavaScript`, `C`, `C++`

## Schema

Each row:

| column | type | contents |
|---|---|---|
| `task` | string | stable task id |
| `lang` | string | English (`en`) or primary programming language |
| `category` | string | detailed recipe category |
| `split` | string | trajectory-disjoint `train` or `val` partition |
| `assistant_step` | int | 1-based target assistant turn |
| `assistant_steps` | int | assistant turns in the source trajectory |
| `target_message_index` | int | index of the final assistant target |
| `n_messages` | int | cumulative message count through the target |
| `messages` | list of objects | cumulative context ending at the target |

`messages` is native JSON.

## Layout

The Hugging Face viewer reads the validated active Parquet shards listed in `dataset-manifest.json`. The equivalent canonical `traces.jsonl` is also published for direct download and conversion. It currently contains
11,235 cumulative next-step rows derived from 2,161 accepted
trajectories over disjoint train and validation tasks.

When training from the cumulative view, supervise only the final assistant
message in each row. Supervising every assistant span would repeatedly
overweight early steps because those spans recur as context in later prefixes.

## Intended use

Supervised fine-tuning of instruction-following, tool-calling, and coding
agents, plus analysis of multi-step planning, parallel calls, tool selection,
state tracking, build-test-fix loops, and verification-driven completion.

## Provenance

Generated with Claude Fable 5 (`anthropic/claude-fable-5`), then filtered by deterministic verification and
explicit Codex acceptance review. Codex supplies no demonstration content; it
only judges or requests replacement of candidate traces. Provider credentials,
user keys, and host-identifying data are scrubbed before publication.

## License

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) — free for training,
research, commercial products, modification, redistribution, and inclusion in
other datasets or corpora, with attribution.

Suggested attribution:

> Claude Fable 5 Agent Traces —
> https://huggingface.co/datasets/greghavens/fable-5-coding-and-debugging-traces