Skip to main content
synth-ai filter reads a trace database and exports the subset of sessions that match your grading rules to SFT-ready JSONL.
  • The command only needs a TOML config; the required [filter] table tells it which trace DB to read, where to write the JSONL, and which thresholds to apply.
  • Config values are validated with the same schema that powers the training tooling, so typos (missing db, bad score thresholds, etc.) are caught up front.
  • Each accepted session becomes a single JSONL record containing user/assistant turns plus metadata (session id, task, seed, rewards, model, timestamps).
  • Judge score thresholds, official reward ranges, split/model filters, and created-at windows are all optional—you can combine them to carve out exactly the dataset you need.
  • When metadata contains structured (multimodal) messages, the export preserves that structure so downstream tooling can still replay images or tool calls.
  • If no sessions match, the command exits politely instead of writing an empty file, so you know to adjust the filters.

​
Options

  • --config PATH — Required. Points to the TOML file containing the [filter] block (db, output, optional thresholds). The CLI refuses to run without it.
uvx synth-ai filter --config configs/filter_local.toml
Example config (configs/filter_local.toml): 
[filter]
db = "traces/v3/synth_ai.db"
output = "ft_data/crafter_high_reward.jsonl"
splits = ["train"]
min_official_score = 0.6
min_judge_scores.qa_accuracy = 0.8
limit = 200
Typical output: 
✓ Config validated: db=traces/v3/synth_ai.db, output=ft_data/crafter_high_reward.jsonl
  → Filtering for official score >= 0.6
  → Limiting to 200 examples
Wrote 148 examples -> /Users/synth/qa/sdk/ft_data/crafter_high_reward.jsonl

​
Notes

  • The command loads the trace DB with SessionTracer, so both SQLite files and Turso URLs are supported (sqlite+aiosqlite:///path/to.db works out of the box).
  • Score helpers respect both minimum and maximum bounds. You can mix min_official_score, max_official_score, min_judge_scores.<name>, and max_judge_scores.<name> in the same config.
  • When judge scores live in trace metadata, the exporter pulls them directly; otherwise the field is left unset in the output JSONL.
  • Multimodal prompts stay intact: user messages containing image blocks or tool directives are written exactly as the task app recorded them.