horus14-wq/fable5-v2-coding-t3

Fable-5 V2 Coding Traces + T3 Transformations

A retrieval-ready corpus of coding and software-engineering reasoning traces, plus three T3-transformed variants of the same episodes, for retrieval-augmented reasoning over agent trajectories.

Ready-to-query: prebuilt t3rag indexes of all four files live at horus14-wq/fable5-v2-coding-t3-indexes — t3rag pull --repo horus14-wq/fable5-v2-coding-t3-indexes --subdir struct and search immediately, no GPU or embedding step needed.

The paper: RAG over Thinking Traces (T3)

RAG over Thinking Traces Can Improve Reasoning Tasks (Negar Arabzadeh, Wenjie Ma, Sewon Min & Matei Zaharia, UC Berkeley; code) argues that retrieval-augmented generation's weak results on reasoning tasks come from the choice of corpus, not from retrieval itself. Instead of retrieving documents, the paper retrieves thinking traces — the intermediate reasoning trajectories a model produces while solving a problem — and introduces T3 (Transformation of Thinking Traces), an offline method that rewrites raw traces into three retrieval-friendly representations: Struct (a concise step-by-step cheatsheet), Semantic (three levels of increasing abstraction down to a single core insight), and Reflect (common failure patterns, misleading intuitions, and critical checks). A plain retrieve-then-generate pipeline over these transformed traces improves reasoning on benchmarks such as AIME 2025–2026, LiveCodeBench, and GPQA-Diamond.

What we did to the original dataset

We applied the T3 recipe to the coding subset of the Fable-5 Premium V2 agent-trace corpus. From ~85k source traces we kept 53,842 software-engineering episodes, normalized each into a lean question + reasoning-trace JSONL record (dropping seed-authoring boilerplate and off-domain rows), then removed 1,094 exact-duplicate episodes and ran all three T3 transformations over the remaining traces with Claude Haiku (claude-haiku-5-5) through the Anthropic API. The paper's prompts were adapted for code: the problem statement is injected into the prompt, the math \boxed{...} answer slot is replaced with a one-line Deliverable, and t3_reflect may return N/A — solution-only trace for clean solutions. Every file — raw and all three variants — contains 52,748 records aligned through source_idx. In the raw file, the upstream trace field was dropped because it was byte-identical to text in all 52,748 records.

Files

File	Rows	Contents
f5_v2_filtered_coding.jsonl	52,748	Raw filtered coding episodes (deduplicated): question, text (reasoning trace), source_idx, n_thoughts, n_actions, source, filter — upstream trace field dropped (byte-identical to text)
f5_v2_t3_struct.jsonl	52,748	T3 Struct cheatsheets in output, plus question, model, variant, usage
f5_v2_t3_reflect.jsonl	52,748	T3 Reflect failure-pattern cards in output
f5_v2_t3_semantic.jsonl	52,748	T3 Semantic three-pass abstractions in output

The three transformed files are aligned row-for-row by source_idx; join them against f5_v2_filtered_coding.jsonl on that key. usage carries the transformation model's token counts.

Intended use

Retrieval corpus for reasoning-RAG over coding trajectories — e.g. indexing each file as a separate corpus and retrieving transformed traces as inference-time guidance, per the T3 paper and the companion t3ragrust retrieval tool.

Source Data & Credit

• Original dataset: saidutta69/fable-5-premium-v2 — Fable-5 Premium V2 agent traces (MIT license)
• Method: T3, from RAG over Thinking Traces Can Improve Reasoning Tasks — official code at github.com/Narabzad/t3
• Transformation model: Claude Haiku (claude-haiku-5-5), Anthropic API

Citation

@misc{arabzadeh2026ragthinkingtracesimprove,
      title={RAG over Thinking Traces Can Improve Reasoning Tasks}, 
      author={Negar Arabzadeh and Wenjie Ma and Sewon Min and Matei Zaharia},
      year={2026},
      eprint={2605.03344},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2605.03344}, 
}