Netflix DPO Reasoner

Upload Dataset

Files are sent to the backend, persisted on this host, and traced in the debug log below.

Drop file or choose one Supports `.json`, `.jsonl`, `.csv`, and plain-text JSONL.

Objective Reasoning style

Output depth Preferred export

Status Waiting for upload

Rows 0

Columns 0

Readiness -

Mode -

Parse -

Trace -

Frontend events plus backend trace references for the current upload.

No events yet.

This UI now assumes contextual ranking first and DPO second.

Each row represents one plus a visible and a . That is the raw contract now.
Fields like and are not optional decoration. They are the real input state that makes ranking personal instead of generic.
The is useful for auditable training. It is not the same thing as hidden chain-of-thought, and it should stay short and reviewable.
Real production logs still need . Without them you can train imitation or ranking, but you cannot honestly claim causal uplift.

Array of objects, newline-delimited JSON objects, or CSV with a header row, including semicolon-delimited CSV.

Weak logic, hidden assumptions, and missing links are called out directly.

Readiness 0 / 100

Detected mode No dataset

These export directly from the current uploaded file and selected options.