Datasets:
chosen listlengths 3 3 | rejected listlengths 3 3 | metadata dict |
|---|---|---|
[
{
"content": "You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corresponding to chapter 29.\n\nBackground you have access to:\n<context>\n{\n \"character\": \"Jim Hawkins\",\n \"intrapersonal_axes\": [\n {\n \"axis_type\": \"intrapersonal\",\n \"axis_id\": \"final_... | [
{
"content": "You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corresponding to chapter 29.\n\nBackground you have access to:\n<context>\n{\n \"character\": \"Jim Hawkins\",\n \"intrapersonal_axes\": [\n {\n \"axis_type\": \"intrapersonal\",\n \"axis_id\": \"final_... | {
"anchor_id": "jim_hawkins_final_intra_01_inworld_a3__p3__arc__gpt-5.4-mini__s0",
"anchor_phase_idx": 3,
"anchor_phase_label": "Mature self-command",
"anchor_typicality": "typical",
"axis_id": "final_intra_01",
"axis_name": "From sheltered dependence to self-directed, relationally anchored agency",
"axis... |
[
{
"content": "You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corresponding to chapter 29.\n\nBackground you have access to:\n<context>\n{\n \"character\": \"Jim Hawkins\",\n \"intrapersonal_axes\": [\n {\n \"axis_type\": \"intrapersonal\",\n \"axis_id\": \"final_... | [
{
"content": "You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corresponding to chapter 29.\n\nBackground you have access to:\n<context>\n{\n \"character\": \"Jim Hawkins\",\n \"intrapersonal_axes\": [\n {\n \"axis_type\": \"intrapersonal\",\n \"axis_id\": \"final_... | {
"anchor_id": "jim_hawkins_final_intra_01_inworld_a3__p3__arc__gpt-5.4-mini__s0",
"anchor_phase_idx": 3,
"anchor_phase_label": "Mature self-command",
"anchor_typicality": "typical",
"axis_id": "final_intra_01",
"axis_name": "From sheltered dependence to self-directed, relationally anchored agency",
"axis... |
[
{
"content": "You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corresponding to chapter 29.\n\nBackground you have access to:\n<context>\n{\n \"character\": \"Jim Hawkins\",\n \"intrapersonal_axes\": [\n {\n \"axis_type\": \"intrapersonal\",\n \"axis_id\": \"final_... | [
{
"content": "You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corresponding to chapter 29.\n\nBackground you have access to:\n<context>\n{\n \"character\": \"Jim Hawkins\",\n \"intrapersonal_axes\": [\n {\n \"axis_type\": \"intrapersonal\",\n \"axis_id\": \"final_... | {
"anchor_id": "jim_hawkins_final_intra_01_inworld_a3__p3__arc__gpt-5.4-mini__s0",
"anchor_phase_idx": 3,
"anchor_phase_label": "Mature self-command",
"anchor_typicality": "typical",
"axis_id": "final_intra_01",
"axis_name": "From sheltered dependence to self-directed, relationally anchored agency",
"axis... |
[
{
"content": "You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corresponding to chapter 29.\n\nBackground you have access to:\n<context>\n{\n \"character\": \"Jim Hawkins\",\n \"intrapersonal_axes\": [\n {\n \"axis_type\": \"intrapersonal\",\n \"axis_id\": \"final_... | [
{
"content": "You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corresponding to chapter 29.\n\nBackground you have access to:\n<context>\n{\n \"character\": \"Jim Hawkins\",\n \"intrapersonal_axes\": [\n {\n \"axis_type\": \"intrapersonal\",\n \"axis_id\": \"final_... | {
"anchor_id": "jim_hawkins_final_intra_01_inworld_a3__p3__arc__gpt-5.4-mini__s1",
"anchor_phase_idx": 3,
"anchor_phase_label": "Mature self-command",
"anchor_typicality": "typical",
"axis_id": "final_intra_01",
"axis_name": "From sheltered dependence to self-directed, relationally anchored agency",
"axis... |
[{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | [{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | {"anchor_id":"jim_hawkins_final_intra_01_inworld_a3__p3__arc__gpt-5.4-mini__s1","anchor_phase_idx":3(...TRUNCATED) |
[{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | [{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | {"anchor_id":"jim_hawkins_final_intra_01_inworld_a3__p3__arc__gpt-5.4-mini__s1","anchor_phase_idx":3(...TRUNCATED) |
[{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | [{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | {"anchor_id":"jim_hawkins_final_intra_01_inworld_a3__p3__arc__gpt-5.4-mini__s2","anchor_phase_idx":3(...TRUNCATED) |
[{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | [{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | {"anchor_id":"jim_hawkins_final_intra_01_inworld_a3__p3__arc__gpt-5.4-mini__s2","anchor_phase_idx":3(...TRUNCATED) |
[{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | [{"content":"You are Jim Hawkins, from \"Treasure Island\". You are at the point in the story corres(...TRUNCATED) | {"anchor_id":"jim_hawkins_final_intra_01_inworld_a3__p3__arc__gpt-5.4-mini__s2","anchor_phase_idx":3(...TRUNCATED) |
[{"content":"You are Anne Shirley, from \"Anne of Green Gables\". You are at the point in the story (...TRUNCATED) | [{"content":"You are Anne Shirley, from \"Anne of Green Gables\". You are at the point in the story (...TRUNCATED) | {"anchor_id":"anne_shirley_final_intra_04_outworld_a2__p2__arc__gpt-5.4-mini__s0","anchor_phase_idx"(...TRUNCATED) |
ArcANE: Do Role-Playing Language Agents Stay in Character at the Right Time?
π Accepted at EMNLP 2026 (Main)
ArcANE is a benchmark for evaluating phase-aware character role-playing. It contains 570 Character Arcs and 4,854 probes across 19 novels and 87 principal characters. Each arc segments a characterβs narrative into phases that capture psychological and behavioral changes, while each probe presents the same scenario across phases to test whether responses evolve consistently.
π Dataset Statistics
| Configuration | Split | Examples |
|---|---|---|
| SFT | train |
20,560 |
| SFT | test |
2,285 |
| DPO | train |
12,746 |
| DPO | test |
375 |
| RLVR | train |
2,513 |
| RLVR | test |
678 |
| Total | train |
35,819 |
| Total | test |
3,338 |
π Novel Coverage
All three configurations cover the same 10 training novels. The two novels East Lynne and The Underdogs are excluded.
| Novel | SFT train | SFT test | DPO train | DPO test | RLVR train | RLVR test |
|---|---|---|---|---|---|---|
Anne of Green Gables |
2,452 | 272 | 1,549 | 44 | 265 | 79 |
Dorian Gray |
1,717 | 191 | 1,007 | 32 | 219 | 63 |
Great Expectations |
1,312 | 146 | 781 | 27 | 167 | 42 |
Jane_Eyre |
2,446 | 272 | 1,538 | 37 | 288 | 81 |
Lady Audley's Secret |
1,912 | 212 | 1,211 | 34 | 237 | 60 |
Little Women |
2,978 | 331 | 1,911 | 57 | 356 | 90 |
Olaudah Equiano |
767 | 85 | 457 | 13 | 106 | 28 |
Persuasion |
2,122 | 236 | 1,318 | 33 | 282 | 76 |
Treasure Island |
1,968 | 219 | 1,186 | 47 | 243 | 76 |
pride_and_prejudice |
2,886 | 321 | 1,788 | 51 | 350 | 83 |
| Total | 20,560 | 2,285 | 12,746 | 375 | 2,513 | 678 |
π Quick Start
from datasets import load_dataset
REPO_ID = "holi-lab/ArcANE-Data"
sft_dataset = load_dataset(REPO_ID, "sft")
dpo_dataset = load_dataset(REPO_ID, "dpo")
rlvr_dataset = load_dataset(REPO_ID, "rl")
# A split can also be loaded directly.
rlvr_train = load_dataset(
REPO_ID,
"rl",
split="train",
)
print(rlvr_train.features)
print(rlvr_train[0])
π§ Data Configuration
A source record is canonical when phase_idx == anchor_phase_idx. Other
phases are counterfactual: the same character answers the same probe from a
different point in their development arc.
ββ canonical phase βββββββ preferred / reference
Novel β Character β Probe ββ€
ββ counterfactual phase ββ plausible off-phase behavior
1. SFT data
Chat-style supervised examples containing:
id: unique generation identifiersource: source novelmeta_data: character, axis, probe, phase, reference, and generation metadatamessages: chat-ready[system, user, assistant]conversation
{
"id": "...",
"source": "...",
"meta_data": {
"character": "...",
"probe_id": "...",
"phase_idx": 0,
"anchor_phase_idx": 0,
"scenario": "...",
"question": "...",
},
"messages": [
{"role": "system", "content": "Character and arc context..."},
{"role": "user", "content": "Scenario and question..."},
{"role": "assistant", "content": "In-character response..."},
],
}
2. DPO data
Conversational preference pairs that contrast canonical behavior with behavior from another phase of the same probe:
chosen: canonical[system, user, assistant]conversationrejected: counterfactual[system, user, assistant]conversationmetadata: probe identity, phase labels, phase distance, and source IDs
The chosen and rejected sequences share the same canonical system/user prompt; only the assistant response differs.
3. RLVR data
Canonical-phase prompts for optimizing one generated response. The assistant turn is omitted from the policy-visible prompt.
data_source: reward-routing identifierprompt: policy-visible[system, user]conversationability:persona_arc_fidelityreward_model: reference style and serialized action/speech/thought targetextra_info: probe identity, phase data, references, and prompt text
Gold action, speech, and thought remain in reward fields only.
π Citation
@misc{song2026arcaneroleplayinglanguageagents,
title={ArcANE: Do Role-Playing Language Agents Stay in Character at the Right Time?},
author={Woojung Song and Nalim Kim and Sangjun Song and Chaewon Heo and Jongwon Lim and Yohan Jo},
year={2026},
eprint={2606.05553},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2606.05553},
}
- Downloads last month
- 55