Skip to main content
rlcli reads chat dumps from various sources and normalizes them into a single messages JSONL format that train sl expects. Use rlcli import to flatten content blocks, normalize roles, filter short threads, and pipe the results directly into training.

​
Usage

​
Positional argument

​
string
required
Path to the input file, or - to read from stdin.

​
Flags

​
choice
default:"messages"
Input format.
  • messages: {"messages": [...]} per line (default).
  • openai: chat-completions dumps. Tool-calls turns are dropped.
  • anthropic: Messages API dumps. Content blocks are flattened into plain text; tool_use, tool_result, image, and thinking blocks are dropped. Top-level system is kept as a system message.
​
string
default:"-"
Output path, or - for stdout (default).
​
int
default:"2"
Minimum number of messages a conversation must contain to be kept.
​
int
Maximum number of conversations to output.

​
Role normalization

rlcli normalizes roles as follows:
  • human → user
  • ai / model → assistant
  • system → system
  • tool, function, tool_result → dropped

​
Content flattening

  • Plain strings are kept as-is.
  • Anthropic-style content blocks: only type: "text" blocks are kept and joined with newlines. All other block types are dropped.

​
Validation

A conversation is kept only when it has at least --min-messages messages and at least one assistant message after normalization.

​
Errors

  • ImportFormatError is raised with a line number on malformed JSON or unknown format strings.
  • If no usable conversations are found, the command exits with code 1.

​
Output

The command writes the normalized JSONL to stdout or --out and prints a count to stderr:

​
Examples

Convert an OpenAI chat-completions dump to a local file:
Pipe imported Anthropic traces straight into supervised fine-tuning: