SMLBuilder/SML-Conversation-Trees

SML Conversation Trees + Python Decoder

Try the live conversation explorer

Explore recorded conversation trees while preserving every branch, the full conversation history and the exact location of each turn in the source JSON.

This public release brings together 1,309 JSON source files from SMLBuilder's existing public corpora, a standalone Python decoder, and a live interactive demo. The catalog includes branching conversations, linear chats and derived training files. It is a source-file count, not a claim that every file is a conversation tree.

Public source datasets

• AAC Children Conversations — pinned at c060d83904330dd578167e29d7714becd7aaec17.
• TinkyBrain TrainingData — pinned at b7f5285c28d77359687c59f2b948b66aefe869c7.

catalog.json points directly to these files at their immutable source revisions. This avoids uploading another copy of the full training corpora. Fields include id, repo, revision, path, bytes, sha256 (LFS objects), blob_id (Git objects), category and title. A null SHA-256 means the Git blob ID is the integrity anchor; the live explorer additionally computes SHA-256 after download.

Use the lightweight decoder

Download tree_decoder.py and a tree JSON from either source dataset. Python 3.9 or newer is enough; there are no package dependencies, model weights, API credentials or GPU requirements.

python3 tree_decoder.py conversation.json --list
python3 tree_decoder.py conversation.json --conversation 0 --path 0,2,1
python3 tree_decoder.py conversation.json --all > conversation-traces.jsonl

--all emits every reachable state in every supported conversation root, including all history, available next replies, terminal status, original tile metadata and RFC 6901 JSON pointers. It never merges identical-looking responses from different branches or invents a continuation. Paths and conversation indices are zero-based. Linear messages files are explicitly labeled and have one recorded next message. Training tables without conversation structure are not interpreted as dialogue.

Supported tree families: Q&A, AAC/continuation, doctor/patient, caller/dispatcher, dispatcher/AAC. The original AAC_cli.py is included unchanged alongside the expanded path-preserving decoder.

Trace format

A trace contains conversation_index, conversation_id, root_pointer, node_pointer, branch_path, depth, schema, history, options and terminal. Each turn carries its original text, role or schema role label, pointer, tile_id when present and unmodified metadata. Web exports add the exact pinned source URL, repository revision and verified content hash. CLI exports identify the input filename and its SHA-256. Missing source roles use neutral schema labels, not an inference about the real speaker.

The starter tree is a byte-identical public example bundled for immediate demo startup. Its original path and hash are recorded in the catalog under 25519a9782d5d013.

Scope and verification

The release contains public conversation data references and decoder code. Raw recovery archives, absolute machine paths, credentials and personal runtime history are not part of this public dataset. Existing upstream dataset terms remain applicable; this release does not add a license to underlying material.

An initial audit walked 13,401 reachable states in 22 representative public files, including 11,159 terminal states. 111,118 displayed history/option pointers were resolved against the original JSON and matched their recorded text. This is a format/trace audit of those samples, not an accuracy or safety certification of the full corpora. Unsupported files and malformed paths return explicit errors.

The live demo replays dataset content. It does not generate answers, assess their truth, or provide professional advice through specialist examples. It uses public read-only source files and does not save visitors' paths as personal conversations.