fable-5-coding-and-debugging-traces-synthetic-corrections

# Model Synthetic Corrections  <div align="center"> <h2>1 TRAJECTORIES · 2 TRAINING ROWS · 16 kB</h2> </div> > Generated by **[moonshiner](https://github.com/greghavens/moonshiner)** — an open harness for > distilling verified instruction-following, tool-use, and agentic coding traces. > **Synthetic Corrections companion dataset.** The original dataset is [greghavens/fable-5-coding-and-debugging-traces](https://huggingface.co/datasets/greghavens/fable-5-coding-and-debugging-traces). These are narrowly, synthetically corrected, independently re-judged traces that never passed in the original dataset. 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 - **Narrowly corrected failed trajectories.** Every row starts from a genuine model session that never passed the primary trace judge. The source reasoning is preserved unchanged; only a minimal, explicitly disclosed synthetic correction is 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 | |---|---:|---:|---:|---| | Refactoring & performance | 1 | 100.0% | 100.0% | Restructure safely and improve measured performance without behavior drift. | ## Languages (current drop) `Bash` ## 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 Everything ships in one data file: `traces.jsonl`. It currently contains 2 cumulative next-step rows derived from 1 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 Source failures were generated with Claude Fable 5 (`anthropic/claude-fable-5`). The configured correction model made only narrowly eligible synthetic repairs, after which deterministic verification and the independent `gpt-5.6-sol` trace judge were run again. Source reasoning is preserved unchanged. 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: > Model Synthetic Corrections — > https://huggingface.co/datasets/greghavens/fable-5-coding-and-debugging-traces-synthetic-corrections