Datasets:

YuBenjamin2004
/

discrete-cot-coding-v5

Modalities:

Formats:

Languages:

Size:

Tags:

chain-of-thought

Libraries:

License:

Dataset card Data Studio Files Files and versions

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access is granted on request. Please state your name and intended use.

Log in or Sign Up to review the conditions and access this dataset content.

Discrete-CoT Coding Reasoning Corpus (v5)

Multi-step coding reasoning traces used in the discrete chain-of-thought / text VQ-VAE study.

corpus_v5.jsonl — one JSON object per generated trace.
14,249 traces over 2,509 problems: 6,562 COMPLETE, 7,687 FAILED (filter on the status field).
Each COMPLETE trace has a steps list; each step carries a think (the reasoning compressed in Design B) and an intermediate answer/code. Trace-level fields include plan, final_code, final_think, fidelity, n_steps, and generation metadata (model, sampling).

This is the source corpus. The per-step VQ-VAE training split is derived from the COMPLETE traces (one row per (qid, trace_idx, step)).

Downloads last month: 13

Total file size: