Skip to main content
Beton
telegram_chat telegram_mvp_v1

Telegram Chat

Simulates a Telegram-like chat environment with user interactions. Different event shape from the PostHog web scenarios — useful for testing agent portability across data sources.

Value metric

Chat engagement / retention

Success event

event_signup

entity_type: user

Scale

  • 0 accounts
  • 0 users/account (mean)
  • 0 sessions/user (mean)
  • 30 days of history

Research metrics proxied

  • — Active conversation rate
  • — Response rate
  • — Retention

Signal paths

Positive paths end in event_signup. Negative paths do not. Every generated event_id belonging to a path is recorded in ground_truth.json.

Positive signals (2)

message_reply_signup ×90
message_sent → reply_received → event_signup
cohorts: high_intent, medium_intent, power_user
reaction_chain_signup ×40
message_sent → reaction_received → event_signup
cohorts: high_intent, power_user

Negative signals (2)

message_then_silence ×80
message_sent → reaction_received
cohorts: low_intent, lurker
leave_after_reply ×55
message_sent → reply_sent → leave_group
cohorts: low_intent, noisy_bot_like

Generate this dataset

Config file: configs/telegram_mvp.yaml

Quickstart 
# Dockerized Postgres (recommended for inspection)
docker compose up -d

uv run dryfit \
  -c configs/telegram_mvp.yaml \
  --dsn postgresql://dryfit_writer:dryfit_writer@127.0.0.1:54329/dryfit \
  --print-summary

# Or local Postgres
./scripts/generate-local -c configs/telegram_mvp.yaml --print-summary

Full setup instructions are in the repo's README — including local Postgres, Grafana inspection, and dataset restore.

Noise parameters

DryFit injects realistic noise on top of the generated signal paths. These probabilities are per-event. Noise never touches rows referenced by ground_truth.json — your scoring logic can trust the truth file is exact.

missing event probability
5.0%
duplicate event probability
2.0%
out of order probability
4.0%
null property probability
3.0%
anonymous actor probability
1.0%
weird property probability
2.0%

Other scenarios

Combined coverage (all models)

→ upgrade_clicked

Union of event types across all business models

Contact / record-based SaaS

→ contact_created

Contacts, leads, subscribers, accounts managed

Credits / token-based

→ credits_purchased

Credits consumed, tokens used, compute units

Event-volume SaaS

→ custom_event_tracked

Events tracked, data points ingested, log lines
Browse all scenarios →

Benchmark your detector against Telegram Chat

Clone the repo, run the config, check your agent's output against ground_truth.json.

View on GitHub