sidechat-gen / README.md
jptaylor's picture
sidechat-gen-0.2: generated conversations as Parquet, with the card
02b7be2 verified
|
Raw History Blame Contribute Delete
4.23 kB
metadata
license: bsd-2-clause
language: en
pretty_name: Sidechat generated conversations
task_categories:
  - text-classification
tags:
  - pipecat
  - addressee-detection
  - multi-party
  - turn-taking
  - synthetic
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: validation
        path: data/validation-*.parquet
      - split: test
        path: data/test-*.parquet

Sidechat generated conversations (sidechat-gen-0.2)

Written conversations between several people and one voice agent, one row per line. Each line is labelled with who it was for and whether the agent should speak. This is all the training data this project made for Pipecat Sidechat.

Code: github.com/pipecat-ai/sidechat

What the models trained on

Build of rc0.3.1 Corpora here
Default gen-pilot, gen-0.1, gen-0.2
agent-1x those, and gen-0.3 and gen-0.4

gen-0.5 is in no released build. Both builds also trained on public data that is not here: meeting speech from AMI (CC BY 4.0), and requests to assistants from SLURP's text (CC BY 4.0) and CLINC150 (CC BY 3.0).

How it was made

A planner lays out each conversation: a setting, two to five people, an agent with a name and a role, and a kind for every line, such as a named request to the agent, an unnamed follow-up or a question to the room. Language models, called o, a, c and b here, write the lines to that plan. The label comes from the kind under the labelling guide (sidechat-guide-0.4), not from the writer.

A committee of three other models checked samples against the guide and agreed with 93% to 99% of the labels of the kinds kept. Lines flagged as not matching their brief are left out. Two of gen-0.5's kinds are left out too (follow_up_later, partner_to_people_after_agent, 3,281 lines): the committee did not agree with their labels.

Names are invented. Training replaces speaker names with anonymous labels ("Speaker 1") on most lines, as a diarizer would.

Splits

Splits never cut a conversation. Conversations written by writer b, or with an agent name or a setting kept out of training, are only in test.

Corpus Conversations Train Validation Test
gen-pilot 599 6,263 656 1,451
gen-0.1 8,095 83,732 8,346 20,405
gen-0.2 11,490 129,798 12,443 16,340
gen-0.3 8,993 102,924 9,629 12,728
gen-0.4 7,296 82,835 8,364 10,671
gen-0.5 5,997 60,800 5,630 13,736
All 466,352 45,068 75,331

Columns

Column
row_id, corpus, conv_id, turn_index where the line comes from; rows of one conversation share conv_id
agent_name, agent_aliases, agent_role the agent's card
speaker, text the line to judge
history up to 8 earlier turns, oldest first: speaker, text, agent (the agent said it)
people the people present
over_agent the line was said while the agent was speaking (null: not known)
alone one person present (null: not known)
to who the line was for: agent, person, group, device, nobody, unknown
to_people names, when to is person
intent direct, indirect, open (to the room), about (about the agent), other
expect respond (speak now), wait (speak only if nobody else does), ignore; null when either is fine
respond target probability that the agent should speak now. A question left open to the room is about 0.5
unclear the line fits the agent and someone else equally
kind, rule the planned kind of line and the guide rule that labels it
setting, register, writer, heldout the scene, its tone, which writer wrote it, and whether that writer is held out
unnamed_agent the agent is never named in this conversation
other_tags any other tags, as JSON