README.md

metadata

language:
  - en
license: mit
task_categories:
  - text-generation
  - conversational
tags:
  - human-collaboration
  - grounding
  - mental-model
  - map-task
  - multi-agent
pretty_name: Almanac
size_categories:
  - 1K<n<10K
configs:
  - config_name: follower_next_action
    data_files:
      - split: train
        path: data/sft/follower_next_action/train.jsonl
      - split: test
        path: data/sft/follower_next_action/test.jsonl
  - config_name: guide_next_action
    data_files:
      - split: train
        path: data/sft/guide_next_action/train.jsonl
      - split: test
        path: data/sft/guide_next_action/test.jsonl
  - config_name: follower_mental_model
    data_files:
      - split: train
        path: data/sft/follower_mental_model/train.jsonl
      - split: test
        path: data/sft/follower_mental_model/test.jsonl
  - config_name: guide_mental_model
    data_files:
      - split: train
        path: data/sft/guide_mental_model/train.jsonl
      - split: test
        path: data/sft/guide_mental_model/test.jsonl
  - config_name: raw_timeline
    data_files:
      - split: train
        path: data/raw_timeline/train.jsonl
      - split: test
        path: data/raw_timeline/test.jsonl
  - config_name: grounding_timeline
    data_files:
      - split: train
        path: data/grounding_timeline/train.jsonl
      - split: test
        path: data/grounding_timeline/test.jsonl

Almanac

Human–human collaborative map-reproduction sessions for studying grounding acts, mental models, and next-action prediction in guide–follower teams.

Dataset contents

Component	Location	Description
Raw sessions	data/raw_sessions/{c1,c2}/{study}/	Per-session guide/follower timelines, full action logs, score boards
SFT splits	data/sft/{task}/train.jsonl, test.jsonl	Chat-format fine-tuning data for 4 prediction tasks
Grounding	data/grounding/{c1,c2}/{study}/	LLM-annotated grounding acts on each timeline step
Flat tables	data/raw_timeline/, data/grounding_timeline/	One row per timeline step; train.jsonl / test.jsonl plus all.jsonl
Metadata	metadata/	Split protocol, study index, grounding manifest

Experimental conditions (canvas visibility)

Sessions are grouped by whether the guide can see the follower’s drawing canvas:

Condition	Canvas visibility	Description
c1	Canvas not visible	The follower’s canvas is hidden from the guide; the guide cannot see the follower’s drawing progress.
c2	Canvas visible	The follower’s canvas is visible to the guide; the guide can see the follower’s drawing progress.

In flat tables and SFT metadata, condition / canvas_condition uses c1 or c2 accordingly.

SFT tasks

1. follower_next_action / guide_next_action — predict the next action (message, draw, erase, undo, reset).
2. follower_mental_model / guide_mental_model — predict mental-model fields at each step.

Each JSONL row includes messages (system / user / assistant), plus metadata: task, role, canvas_condition, study_name, participant_name, timeline_index, step_index.

Grounding annotations

Grounding labels (grounding_act, grounding_rationale) were generated with Azure OpenAI gpt-5.5 using the project prompt in Grounding Acts Human Annotation/llm_propmt.txt. See metadata/grounding_manifest.json for per-study file paths and metadata/studies_index.json for action counts.

Train / test split (by study)

Splits are defined at the session (study) level, not by random steps.

Train — c1: study3, study9, study12, study14, study18, study24, study27, study30, study31
Train — c2: study2, study4, study6, study10, study11, study13, study20, study22, study35, study37

Test — c1: study8, study21, study25
Test — c2: study15, study17, study36

Full definition: metadata/split_protocol.json.

Usage

from datasets import load_dataset

# SFT example
ds = load_dataset("NEU-HAI/Almanac", "follower_next_action", split="train")
print(ds[0]["messages"])

# Flat grounding timeline (study-level train/test splits)
g = load_dataset("NEU-HAI/Almanac", "grounding_timeline", split="test")
print(g[0]["grounding_act"], g[0]["action_content"])

Regenerating this bundle

From the repository root:

python Almanac/scripts/prepare_almanac.py

Upload to Hugging Face Hub

hf auth login
cd Almanac
hf upload NEU-HAI/Almanac . --repo-type=dataset

For large updates or resumable uploads:

hf upload-large-folder NEU-HAI/Almanac . --repo-type=dataset

License

This dataset is released under the MIT License.

Citation

If you use Almanac, please cite the associated EMNLP 2026 paper (bibtex to be added).