Skip to content

Repository files navigation

OFL Evals

Evaluation frameworks for AI-assisted facilitation — measuring facilitation quality through structured comparison rather than subjective assessment.

Approach

Instead of asking "Is this good facilitation?" (subjective), OFL evals ask comparative questions:

  • "How similar is this conversation to Socratic dialogue?"
  • "How close is this facilitation to restorative justice principles?"

This relative comparison produces measurable feedback signals for training and improving AI facilitators.

The Why-How-Who Framework

Every facilitation conversation can be characterized along three dimensions:

Why (Purpose & Outcomes)

Outcome Type Description Example Methods
Agreement Building Help group reach consensus Delphi, Consensus Workshop
Preference Elicitation Surface individual preferences Polling, Harmonica
Error Surfacing Identify gaps in thinking Devil's Advocate, Red Team
Perspective Taking Expose to other viewpoints Cross-pollination, Fishbowl
Synthesis Aggregate into actionable output Affinity Mapping
Ideation Generate new ideas Brainstorming, Six Hats
Conflict Resolution Resolve disagreements NVC, Mediation

How (Process & Techniques)

Intervention styles: Non-directive (logistics only) → Semi-directive (questions, summaries) → Directive (guide thinking, challenge)

Question types: Open, Closed, Probing, Clarifying, Challenging

Timing: Scheduled (predetermined) / Responsive (react to input) / Threshold-based (conditions met)

Who (Participants & Dynamics)

Interaction modes: One-to-one, Small group (3-12), Large group (12+), Plenary

Power dynamics: Hierarchy sensitivity, anonymity support, minority voice protection

Conversation Signatures

The key mechanism: compute "signatures" of conversations based on Why-How-Who dimensions, then compare them to reference methodologies.

  1. Label dialogue acts with Why-How-Who tags
  2. Count frequencies of each tag type
  3. Create a vector representing conversation characteristics
  4. Compare to known methodology signatures (distance = similarity measure)

Evaluation Metrics

Category Metrics
Process Intervention frequency/timing, question type distribution, speaking time balance, topic coverage
Outcome Agreement level, idea quantity/quality, participant satisfaction
Signature Distance to target methodology, consistency within session, appropriate adaptation

Transcript Processor

A CLI tool that processes raw facilitation transcripts into anonymized, WHoW-annotated benchmarks.

npm install
npx tsx src/process-transcript.ts --input <file> --methodology <name> --date <YYYY-MM-DD> [--facilitator <names>] [--dry-run]

Pipeline: parse (auto-detect format) → anonymize (names, PII) → annotate (gpt-4o-mini WHoW tagging) → signature (tag frequency vectors) → render (markdown + YAML frontmatter).

See src/ for implementation. Add your OpenAI key to .env.

Judge Calibration

Signatures are comparative, but some quality questions still reduce to a judgment ("is this a good close?"). Where an LLM-as-judge scores such an axis, it needs grounding in expert taste — the gap the weval methodology addresses. calibration/ holds real, anonymized facilitation turns + the rubric criterion in plain English, for expert facilitators to label good / weak. Those labels recalibrate the judge so it scores like an expert, not a model default. First pack: the closing-turn axis.

Repository Structure

why-how-who-framework.md    # Full framework specification
src/                        # Transcript processor CLI
schemas/                    # Data schemas for evaluation results
prompts/                    # LLM prompts for automated tagging
benchmarks/                 # Annotated benchmark transcripts
transcripts/                # Raw input transcripts
calibration/                # Human labels calibrating LLM-as-judge to expert taste

Applications

  • Pattern development — Encode and compare facilitation methods using consistent dimensions
  • AI training — Generate labeled datasets and feedback signals for reinforcement learning
  • Quality assessment — Automated comparison of conversations to reference methodologies

Related Work

Fora Corpus (MIT, ACL 2024)

Fora is a corpus of 262 facilitated dialogues (39,911 turns) with human annotations for 7 facilitation strategies and 2 personal sharing types. It's the closest academic dataset to what OFL evals produces. Key differences:

  • Fora annotates facilitator strategies (follow-up questions, making connections, etc.) and participant sharing — narrower but human-validated
  • WHoW annotates purpose (why), technique (how), and interaction direction (who) — broader, captures dimensions Fora doesn't (purpose, directionality)
  • Complementary: running WHoW annotation on Fora transcripts would validate our taxonomy against human ground truth

Fora data is available by request from github.com/schropes/fora-corpus.

ConvoKit (Cornell)

ConvoKit is a Python toolkit for conversational analysis with a standardized corpus format and built-in analysis transformers. Used by 30+ datasets including Fora.

Relevant analysis tools that would work on our data if we export in ConvoKit format:

  • Politeness Strategies — lexical/parse-based politeness scoring (facilitator tone measurement)
  • Linguistic Coordination — measures whether participants mirror facilitator language (rapport signal)
  • CRAFT Forecasting — predicts conversation derailment (evaluates whether facilitation prevents breakdown)
  • Redirection detection — built-in tool for what our redirect tag captures
  • Linguistic Diversity — quantifies speaker variation within conversations

Our pipeline handles the upstream problem ConvoKit doesn't: raw transcript → structured annotated corpus. ConvoKit handles the downstream analysis. Adding ConvoKit-format export would make all these tools available on our benchmarks.

weval (Collective Intelligence Project)

weval is an open, CC0 platform for public AI evals: evals are written as blueprints — YAML pairing prompts with should/should_not rubrics, scored by LLM-as-judge with multi-judge consensus and inter-rater reliability (Krippendorff's α). Two ways it connects to OFL evals:

  • Format (interop target). A Why-How-Who facilitation eval can be expressed as a weval blueprint — making OFL evals runnable on weval-shaped infrastructure, not just this pipeline. Planned interop direction (the eval-side analog of a ConvoKit export).
  • Judge calibration. weval's methodology — consensus judging + Krippendorff's α reliability bands + bias countermeasures — is prior art for grounding any LLM-as-judge scoring added here.

CIP run weval and have been a proposed validation partner for the OFL eval suite. See the weval wiki page and the blueprint format.

Other References

License

MIT

About

Evaluation frameworks for AI-assisted facilitation — Why-How-Who conversation signatures

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages