ai4he/llm-judge-teammate-coding

LLM-Judge Teammate Coding Scheme

Dataset accompanying the paper LLM-Judge Behavioral Coding Scheme for Agent Teammate Quality (14th International Conference on Human-Agent Interaction (HAI '26), 2026).

The full stimulus set (single-turn, matched-register and subtle-stress partner lines; multi-turn dialogues are defined in the runner script), the exact judge system prompt, 1,338 judge rating vectors from five LLM judges across three stimulus families, the human pilot (ratings, results, blinding key and the blank rating instrument), the camera-ready summary statistics, and the stimulus, runner and analysis code.

Paper	LLM-Judge Behavioral Coding Scheme for Agent Teammate Quality
Venue	14th International Conference on Human-Agent Interaction (HAI '26), 2026
Authors	Carlos Toxtli, Manuel Delaflor
License	Creative Commons Attribution 4.0 International (cc-by-4.0); code: MIT License
Dataset DOI	10.5281/zenodo.24179034
Version	1.0.0
Maintainer	Human-AI Empowerment Lab, Clemson University

Paper abstract

Evaluating whether an AI agent behaves like a good teammate is becoming a bottleneck as multi-agent language-model systems proliferate, because human ratings do not scale to the volume of transcripts these systems generate. We ask whether a language model can stand in for a human rater of teammate quality, and we present a controlled proof-of-concept validation of a measurement instrument for that purpose rather than a large-scale study. Drawing on the autonomous-agent teammate-likeness construct from the human-autonomy-teaming literature, we define a six-dimension behavioral coding scheme, covering altruistic, benevolent, interdependent, emotive, communicative, and synchronized conduct, and adapt it for use by a language-model judge. We validate the adaptation with experiments in which the partner's lines in team-task transcripts are rewritten as cooperative, neutral, or defective while the rest of the conversation is held fixed. Five judges from five model families (GLM, GPT-OSS, Qwen, Gemma, and Mistral) rate single-turn scenarios and multi-turn dialogues, and a surface-feature audit with length- and politeness-matched registers tests whether the judges read teammate conduct rather than verbosity. Every judge orders the manipulated behavior as intended, most with fully separated confidence intervals between adjacent conditions, and every pair of judges agrees strongly on relative ordering, though absolute calibration differs across judges by up to about a scale point in the middle of the range. On matched registers the judges still separate cooperative from degraded conduct, but resolution at the subtle-defective boundary is limited. A blinded single-rater human pilot on all stimuli corroborates both findings: the rater closely reproduces the judges' orderings and shows the same resolution limit at the subtle-defective boundary. The scheme is therefore valid for relative comparisons, such as ranking systems or versions, but not yet for absolute teammate-quality thresholds applied across judges, and not yet for fine distinctions among subtly poor teammates. We give design implications for building an LLM-judge teaming-evaluation pipeline, and release the scheme, stimuli, judge prompts, and all individual ratings from this version's experiments.

Dataset structure

Path	Format	Size	Records
code/_common/llm_client.py	py	10.5 KB
code/analysis/analyze_camera_ready.py	py	12.3 KB
code/analysis/analyze_human_pilot.py	py	6.0 KB
code/analysis/analyze_scenario_level.py	py	11.0 KB
code/analysis/compute_demand_baselines.py	py	12.8 KB
code/analysis/make_camera_ready_figures.py	py	10.0 KB
code/analysis/run_psychometric.py	py	5.2 KB
code/human_pilot/xlsx_to_tidy.py	py	1.7 KB
code/stimuli/run_expanded_scenarios.py	py	19.4 KB
code/stimuli/run_new_judges.py	py	5.6 KB
code/stimuli/run_proper_experiment.py	py	6.7 KB
code/stimuli/run_subtle_stress.py	py	15.5 KB
data/human_pilot/blinding_key.csv	csv	2.0 KB	84
data/human_pilot/human_pilot_responses.csv	csv	17.5 KB	504
data/human_pilot/human_pilot_results.json	json	3.9 KB
data/judge_ratings/p049_expanded_multiturn.json	json	28.2 KB	72
data/judge_ratings/p049_expanded_singleturn.json	json	98.4 KB	252
data/judge_ratings/p049_experiment_rerun_aug2026.json	json	36.2 KB	108
data/judge_ratings/p049_newjudge_gemma-4-31b_multiturn.json	json	17.8 KB	36
data/judge_ratings/p049_newjudge_gemma-4-31b_singleturn.json	json	62.1 KB	126
data/judge_ratings/p049_newjudge_gemma-4-31b_subtle.json	json	39.8 KB	84
data/judge_ratings/p049_newjudge_leanstral-2603_multiturn.json	json	14.4 KB	36
data/judge_ratings/p049_newjudge_leanstral-2603_singleturn.json	json	50.6 KB	126
data/judge_ratings/p049_newjudge_leanstral-2603_subtle.json	json	32.0 KB	84
data/judge_ratings/p049_newjudge_qwen3.6-35b-a3b_multiturn.json	json	18.0 KB	36
data/judge_ratings/p049_newjudge_qwen3.6-35b-a3b_singleturn.json	json	62.6 KB	126
data/judge_ratings/p049_newjudge_qwen3.6-35b-a3b_subtle.json	json	36.7 KB	84
data/judge_ratings/p049_subtle_stress.json	json	62.4 KB	168
data/prompts/judge_system_prompt.txt	txt	597 B
data/results/camera_ready_summary.json	json	28.0 KB
data/stimuli/stimuli_singleturn_and_matched_register.csv	csv	22.0 KB	84
docs/README.md	md	3.1 KB
docs/human_pilot/P049_teammate_rating_blank_instrument.xlsx	xlsx	16.6 KB

33 files, 769.4 KB in total (metadata files excluded). Records are rows for CSV/TSV/JSONL files and list length for JSON arrays.

Loading the data

from datasets import load_dataset

ds = load_dataset("ai4he/llm-judge-teammate-coding", "stimuli_singleturn_and_matched_register")

Available subsets: stimuli_singleturn_and_matched_register, human_pilot_responses, blinding_key. All other files can be downloaded with huggingface_hub.snapshot_download or from the Zenodo record.

Known limitations

• The original 108-rating run behind Table 2 was overwritten by the August 2026 re-run, which is the version included.
• Only the last 200 characters of judge reasoning were kept for judges C-E.
• Scripts expect data/ and outputs/ next to them; adjust paths before re-running.

Personal and sensitive information

The human pilot has a single rater (one of the authors), identified only as rater_1; the blank instrument lists an author's public work email.

License

The data and documentation are released under the Creative Commons Attribution 4.0 International. Source code files (.ipynb, .js, .m, .py, .sh) are released under the MIT License.

Citation

If you use this dataset, please cite the paper:

@inproceedings{Toxtli2026LLMJudge,
 title = {LLM-Judge Behavioral Coding Scheme for Agent Teammate Quality},
 author = {Carlos Toxtli and Manuel Delaflor},
 booktitle = {Proceedings of the 14th International Conference on Human-Agent Interaction (HAI ’26)},
 year = {2026},
 url = {https://hai-conference.net/hai2026/program-schedule/}
}

and the dataset:

@dataset{Toxtli2026LlmJudgeTeammateCodingData,
  author = {Toxtli, Carlos and Delaflor, Manuel},
  title = {{LLM-Judge Teammate Coding Scheme}: Dataset for ``LLM-Judge Behavioral Coding Scheme for Agent Teammate Quality''},
  year = {2026},
  version = {1.0.0},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.24179034},
  url = {https://doi.org/10.5281/zenodo.24179034}
}

Further documentation: docs/README.md.