alibaba / alibaba/ROCK

docs: add trajectory distillation best practice guide

Open
#916 0 comments 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
485
Forks
81
Avg merge
16h 12m
Merged PRs (30d)
8

Description

## Summary

Add a user guide documenting the best practice for using the ROCK Job system to collect Agent trajectories for model distillation training.

## Motivation

Trajectory distillation (using a strong Teacher model's behavioral data to train a weaker Student model) is a common use case for the Job system, but there is currently no documentation covering this workflow end-to-end.

## Scope

- User guide document (Chinese + English) under `docs/versioned_docs/version-1.7.x/User Guides/`
- Example code and config template under `examples/trajectory_distillation/`
- Covers: quick start (end-to-end runnable example), configuration details, trajectory data reference (result.json + trajectory.json), advanced usage (async mode, rejection sampling, DPO pairs)
- Validated against a real ROCK deployment with swe-agent on SWE-bench

Contributor guide

No contributing guide indexed for this repository

Research direction

Start by reviewing the existing guides under docs/versioned_docs/version-1.7.x/User Guides/ and the examples directory before adding the trajectory_distillation materials. Done means a Chinese and English guide, runnable example and configuration template covering the listed workflow, data files, and advanced modes, validated against a ROCK deployment with swe-agent on SWE-bench.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
documentation, machine-learning
Issue type
Documentation
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
48/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.