Add trajectory condensation utility for OpenHands SFT data

[neubig]

Summary

This PR adds a utility script that applies OpenHands SDK context condensation to SFT trajectories, splitting them when condensation occurs.

Motivation

Long agent trajectories can exceed context limits during training. The OpenHands SDK includes a context condenser that summarizes and removes older conversation turns when a threshold is exceeded. This utility applies that condensation to stored trajectories, creating training data that reflects the condensed context that would be seen during inference.

Implementation

Main Components

1. condense_trajectories.py: Main utility script that:

   • Loads SFT trajectories from JSON
   • Converts conversations to MessageEvent objects
   • Applies context condensation using OpenHands SDK
   • Splits trajectories when condensation occurs (since prefix changes)
   • Outputs condensed trajectories in the same SFT format

2. mock_condenser.py: A mock condenser for testing that:

   • Implements the same interface as LLMSummarizingCondenser
   • Triggers condensation based on simple threshold
   • Creates mock summaries without requiring LLM calls

3. README_condense_trajectories.md: Comprehensive documentation

Key Features

• Flexible condenser support: Can use either LLM-based or mock condenser
• Configurable parameters: max_size and keep_first control condensation behavior
• Trajectory splitting: Creates new trajectory segments when condensation occurs
• Comprehensive logging: Tracks condensation events and segment creation
• Format preservation: Output maintains the same SFT format as input

Input/Output

• Input: sample_sft_openhands.json with N trajectories
• Output: Same format with N*M trajectories, where M = average condensations + 1
• Each segment gets a unique ID: {original_id}_seg{index}

Testing

The script has been tested on multiple datasets:

• agenttuning_alfworld: 5 trajectories → 23 segments (avg 4.6x split)
• swe-smith: 5 trajectories → 29 segments (avg 5.8x split)
• swe-gym: 5 trajectories → 21 segments (avg 4.2x split)

Example verification with jq:

jq '[.[] | {id, conversations: (.conversations | length)}]' output.json

Usage

With mock condenser (for testing):

python scripts/condense_trajectories.py \
  datasets/swe-smith/sample_sft/sample_sft_openhands.json \
  output_condensed.json \
  --max-size 12 \
  --keep-first 2 \
  --use-mock-condenser

With LLM condenser:

export LLM_API_KEY="your-api-key"
python scripts/condense_trajectories.py \
  datasets/swe-smith/sample_sft/sample_sft_openhands.json \
  output_condensed.json \
  --max-size 120 \
  --keep-first 4

Dependencies

• Requires openhands-sdk to be installed
• No changes to existing code or dependencies

Future Work

Potential enhancements:

• Batch processing of entire datasets
• Configuration files for different condensation strategies
• Integration with training pipeline
• Analysis tools for condensation patterns

@neubig can click here to continue refining the PR

[openhands-ai]

Looks like there are a few issues preventing this PR from being merged!

• GitHub Actions are failing:
  • Pre-commit Checks

If you'd like me to help, just leave a comment, like

@OpenHands please fix the failing actions on PR #157 at branch `add-trajectory-condenser-utility`

Feel free to include any additional details that might help me get this PR into a better state.

_{^{You can manage your notification settings}}