Datasets:

Modalities:
Text
Formats:
parquet
License:
nikhilsk commited on
Commit
afca6fb
·
verified ·
1 Parent(s): 2b5efa4

Clarify release documentation and code comments

Browse files
DATA.md CHANGED
@@ -7,8 +7,8 @@
7
  | `evaluation` | `case_id`, `question_id`, `topic`, `query_language`, `query_type`, `query`, `task`, `document_ids`, `target_narrative_id`, `expected_document_ids`, `system_instruction`, `query_context`, `prior_answer`, `followup_instruction`, `input_sha256` |
8
  | `responses` | `model_id`, `case_id`, `response`, `response_sha256`, `language_status`, `score` |
9
 
10
- `document_ids` joins evaluation cases to articles; `narrative_id` identifies the source-perspective and publication-language group. `expected_document_ids` defines scoring targets. Question IDs join translated queries, and case IDs join model responses to evaluation inputs.
11
 
12
- `query_type` is neutral or confirmatory. `task` is NI, NE, MNI, BP or DP. MNI includes its fixed `prior_answer` and `followup_instruction`; `query_context` contains any additional query context. Input and response checksums verify exact evaluation inputs and answers.
13
 
14
- `language_status` is `same`, `different`, `mixed` or `no_answer`; `score` ranges from zero to one. `text_en` contains English text where available. Empty English text or query translations indicate an unavailable version.
 
7
  | `evaluation` | `case_id`, `question_id`, `topic`, `query_language`, `query_type`, `query`, `task`, `document_ids`, `target_narrative_id`, `expected_document_ids`, `system_instruction`, `query_context`, `prior_answer`, `followup_instruction`, `input_sha256` |
8
  | `responses` | `model_id`, `case_id`, `response`, `response_sha256`, `language_status`, `score` |
9
 
10
+ `document_ids` identifies the articles supplied in each evaluation case, and `narrative_id` groups articles by source perspective and publication language. `expected_document_ids` identifies articles associated with the narratives required by the task. Use `question_id` to match queries across languages and `case_id` to match responses to their evaluation inputs.
11
 
12
+ `query_type` distinguishes neutral and confirmatory questions, while `task` identifies Narrative Inclusion (NI), Narrative Exclusion (NE), Missing Narrative Identification (MNI), Baseline Prompt (BP) or Diverse Prompt (DP). MNI includes its fixed `prior_answer` and `followup_instruction`; `query_context` supplies background provided with the question. Input and response checksums verify that evaluation inputs and answers are unchanged.
13
 
14
+ `language_status` records whether a response passed the language check (`same`), failed it (`different` or `mixed`), or contained no answer (`no_answer`); `score` ranges from zero to one. `text_en` contains article text in English; empty values in `text_en` or a `query_<language code>` column indicate that the corresponding English text or query translation is unavailable.
EVALUATION.md CHANGED
@@ -12,7 +12,7 @@ Choose a configuration from `models.json`. Each exported record contains the mes
12
  {"case_id":"case identifier from the input","model_id":"your model identifier","response":"returned answer"}
13
  ```
14
 
15
- MNI includes the fixed prior answer and follow-up. Add `--limit 5` to export a small trial.
16
 
17
  ## Set up language detection
18
 
@@ -44,7 +44,7 @@ python code/narrativebench.py score --predictions predictions.jsonl \
44
 
45
  The command detects response language and computes narrative scores automatically. Low-confidence detections pass the language check and proceed to citation scoring; responses confidently identified as outside the query language receive a score of 0. Language detection and citation scoring run locally, with no LLM calls or manual review.
46
 
47
- NI and NE use narrative F1, while MNI, BP and DP use narrative recall. Results include a score for every response and a summary by model and task. The command also saves the language outcomes in `scores.languages.jsonl`.
48
 
49
  ## Reproduce the recorded scores
50
 
 
12
  {"case_id":"case identifier from the input","model_id":"your model identifier","response":"returned answer"}
13
  ```
14
 
15
+ Missing Narrative Identification (MNI) includes the fixed prior answer and follow-up; add `--limit 5` to export a small trial.
16
 
17
  ## Set up language detection
18
 
 
44
 
45
  The command detects response language and computes narrative scores automatically. Low-confidence detections pass the language check and proceed to citation scoring; responses confidently identified as outside the query language receive a score of 0. Language detection and citation scoring run locally, with no LLM calls or manual review.
46
 
47
+ Narrative Inclusion (NI) and Narrative Exclusion (NE) use narrative F1, while MNI, Baseline Prompt (BP) and Diverse Prompt (DP) use narrative recall. Results include a score for every response and a summary by model and task. The command also saves the language outcomes in `scores.languages.jsonl`.
48
 
49
  ## Reproduce the recorded scores
50
 
LICENSE_NOTICE.md CHANGED
@@ -1 +1 @@
1
- Dataset annotation and curation metadata retain CC BY-NC-SA 4.0. Source-article rights remain with their publishers; article URLs and attribution are provided in the data. Existing upstream software licenses continue to apply.
 
1
+ Dataset annotation and curation metadata retain CC BY-NC-SA 4.0. Source-article rights remain with their publishers; article URLs and attribution are provided in the data. Software dependencies and model weights retain their own licenses.
README.md CHANGED
@@ -48,7 +48,7 @@ configs:
48
 
49
  NarrativeBench evaluates the information seeking capabilities models need to help people navigate diverse narratives across languages. The dataset accompanies *The Narrative Gap: Can LLMs Help Us Navigate Diverse Narratives Across Languages?*
50
 
51
- **[Benchmark code](https://github.com/isle-dev/Narrative-Bench-NeurIPS-2026)** · **[Dataset](https://huggingface.co/datasets/nikhilsk/narrative-bench)**
52
 
53
  ## Contents
54
 
@@ -65,7 +65,7 @@ NarrativeBench evaluates the information seeking capabilities models need to hel
65
 
66
  Articles were gathered through localized searches and grouped by source perspective and publication language. The query set combines 283 directly sourced Reddit questions with 177 topic-specific adaptations, comprising 230 neutral and 230 confirmatory queries.
67
 
68
- Evaluation covers narrative inclusion (NI), narrative exclusion (NE), missing narrative identification (MNI), baseline prompting (BP), and diverse prompting (DP). Each selected narrative contributes three articles to a case. Models answer in the query language and cite the supplied document IDs. NI and NE use narrative F1; MNI, BP and DP use narrative recall. Scores range from zero to one. Automatic language checks use Lingua and fastText; low-confidence detections pass the language check, while responses confidently identified as outside the query language receive zero.
69
 
70
  ## Load and evaluate
71
 
@@ -78,7 +78,7 @@ cases = load_dataset("nikhilsk/narrative-bench", "evaluation", split="test")
78
  responses = load_dataset("nikhilsk/narrative-bench", "responses", split="test")
79
  ```
80
 
81
- Install the `datasets` package to use this example. The [GitHub repository](https://github.com/isle-dev/Narrative-Bench-NeurIPS-2026) provides instructions for running the benchmark, scoring new answers and reproducing the recorded scores.
82
 
83
  ## Citation
84
 
 
48
 
49
  NarrativeBench evaluates the information seeking capabilities models need to help people navigate diverse narratives across languages. The dataset accompanies *The Narrative Gap: Can LLMs Help Us Navigate Diverse Narratives Across Languages?*
50
 
51
+ **[GitHub](https://github.com/isle-dev/Narrative-Bench-NeurIPS-2026)** · **[Dataset](https://huggingface.co/datasets/nikhilsk/narrative-bench)**
52
 
53
  ## Contents
54
 
 
65
 
66
  Articles were gathered through localized searches and grouped by source perspective and publication language. The query set combines 283 directly sourced Reddit questions with 177 topic-specific adaptations, comprising 230 neutral and 230 confirmatory queries.
67
 
68
+ Evaluation covers Narrative Inclusion (NI), Narrative Exclusion (NE), Missing Narrative Identification (MNI), Baseline Prompt (BP), and Diverse Prompt (DP). Each selected narrative contributes three articles to a case. Models answer in the query language and cite the supplied document IDs. NI and NE use narrative F1; MNI, BP and DP use narrative recall. Scores range from zero to one, and automatic language checks use Lingua and fastText; low-confidence detections pass the language check, while responses confidently identified as outside the query language receive zero.
69
 
70
  ## Load and evaluate
71
 
 
78
  responses = load_dataset("nikhilsk/narrative-bench", "responses", split="test")
79
  ```
80
 
81
+ Install the `datasets` package to use this example. The [evaluation guide](EVALUATION.md) explains how to run the benchmark, score new answers and reproduce the recorded scores.
82
 
83
  ## Citation
84
 
code/language_detection.py CHANGED
@@ -13,7 +13,7 @@ LID_CODES={'amh_Ethi':'am','tir_Ethi':'ti','npi_Deva':'ne','prs_Arab':'prs','pes
13
  'spa_Latn':'es','por_Latn':'pt','pol_Latn':'pl','tam_Taml':'ta','tha_Thai':'th','tur_Latn':'tr'}
14
 
15
  class FastTextFallback:
16
- """Pinned local model, loaded without remote code or text uploads."""
17
  def __init__(self,path,manifest):
18
  config=json.loads(Path(manifest).read_text());path=Path(path)
19
  for key in ['probability_threshold','margin_threshold']:
@@ -28,8 +28,7 @@ class FastTextFallback:
28
  required={'__label__amh_Ethi','__label__tir_Ethi','__label__npi_Deva','__label__prs_Arab','__label__pes_Arab'}
29
  if not required<=set(self.model.get_labels()):raise ValueError('language model lacks required classes')
30
  def candidates(self,text):
31
- # Native API avoids the NumPy-2 array-copy incompatibility in the old
32
- # fastText Python predict wrapper; prediction itself is unchanged.
33
  result=self.model.f.predict(text.replace('\n',' ')+'\n',10,0.0,'strict')
34
  return [(label.removeprefix('__label__'),float(prob)) for prob,label in result]
35
 
@@ -57,8 +56,7 @@ def clean(s):
57
  s=re.sub(r'<think>[\s\S]*?</think>',' ',s,flags=re.I)
58
  s=re.sub(r'```[\s\S]*?```',' ',s);s=re.sub(r'^\s*>.*$',' ',s,flags=re.M)
59
  s=re.sub(r'https?://\S+',' ',s)
60
- # Citation bundles/full-width/bare IDs are not language evidence. Repeated
61
- # Dxxxx tokens otherwise cause spurious Latin/Cyrillic classifications.
62
  s=re.sub(r'(?<![A-Za-z0-9_])D\d{4,}(?![A-Za-z0-9_])',' ',s)
63
  return re.sub(r'\s+',' ',s).strip()
64
  def classify(detector,text,query,fallback=None):
 
13
  'spa_Latn':'es','por_Latn':'pt','pol_Latn':'pl','tam_Taml':'ta','tha_Thai':'th','tur_Latn':'tr'}
14
 
15
  class FastTextFallback:
16
+ """Load the fastText language model and verify its file checksum."""
17
  def __init__(self,path,manifest):
18
  config=json.loads(Path(manifest).read_text());path=Path(path)
19
  for key in ['probability_threshold','margin_threshold']:
 
28
  required={'__label__amh_Ethi','__label__tir_Ethi','__label__npi_Deva','__label__prs_Arab','__label__pes_Arab'}
29
  if not required<=set(self.model.get_labels()):raise ValueError('language model lacks required classes')
30
  def candidates(self,text):
31
+ # Use the native predictor for NumPy 2 compatibility.
 
32
  result=self.model.f.predict(text.replace('\n',' ')+'\n',10,0.0,'strict')
33
  return [(label.removeprefix('__label__'),float(prob)) for prob,label in result]
34
 
 
56
  s=re.sub(r'<think>[\s\S]*?</think>',' ',s,flags=re.I)
57
  s=re.sub(r'```[\s\S]*?```',' ',s);s=re.sub(r'^\s*>.*$',' ',s,flags=re.M)
58
  s=re.sub(r'https?://\S+',' ',s)
59
+ # Exclude document IDs from language detection.
 
60
  s=re.sub(r'(?<![A-Za-z0-9_])D\d{4,}(?![A-Za-z0-9_])',' ',s)
61
  return re.sub(r'\s+',' ',s).strip()
62
  def classify(detector,text,query,fallback=None):
code/narrativebench.py CHANGED
@@ -157,7 +157,7 @@ def validate(root):
157
  return counts|{'input_hashes_verified':len(cases),'network_calls':0}
158
 
159
  def provider_payload(conversation,configuration):
160
- """Prepare provider request fields without credentials or network access."""
161
  from urllib.parse import quote
162
  provider=configuration['provider'];model=configuration['api_model'];generation=configuration['generation']
163
  tokens=generation['max_output_tokens'];temperature=generation.get('temperature',0)
@@ -175,8 +175,7 @@ def provider_payload(conversation,configuration):
175
  if temperature is not None:body['temperature']=temperature
176
  return {'url':'https://api.openai.com/v1/responses','body':body}
177
  if provider=='moonshot':
178
- # Kimi's evaluated nonthinking condition has a fixed temperature of0.6.
179
- # Its batch payload omits the fixed parameter and explicitly disables thinking.
180
  body={'model':model,'messages':conversation,'max_tokens':tokens}
181
  body.update(configuration['request_options'])
182
  return {'url':'https://api.moonshot.cn/v1/chat/completions','body':body}
 
157
  return counts|{'input_hashes_verified':len(cases),'network_calls':0}
158
 
159
  def provider_payload(conversation,configuration):
160
+ """Build messages and generation settings for the model provider."""
161
  from urllib.parse import quote
162
  provider=configuration['provider'];model=configuration['api_model'];generation=configuration['generation']
163
  tokens=generation['max_output_tokens'];temperature=generation.get('temperature',0)
 
175
  if temperature is not None:body['temperature']=temperature
176
  return {'url':'https://api.openai.com/v1/responses','body':body}
177
  if provider=='moonshot':
178
+ # Kimi's request format uses a fixed temperature and disables thinking.
 
179
  body={'model':model,'messages':conversation,'max_tokens':tokens}
180
  body.update(configuration['request_options'])
181
  return {'url':'https://api.moonshot.cn/v1/chat/completions','body':body}
manifest.json CHANGED
@@ -10,13 +10,13 @@
10
  "files": {
11
  ".gitattributes": "30129f61f2a1b984b358279fb24769ff73836dc7f839a765b67775ae2af04a13",
12
  ".gitignore": "6843cd7b533a96ccb653dd16ece08e028ce16018b537283a1bfa5354c917f028",
13
- "DATA.md": "13b53cafcb4d4bf8614e2d8f412d70ca4e5214f6fa738642e4e413e257f8f680",
14
- "EVALUATION.md": "99fe5361872761ca8881e36de485740112d2557ce9fe2d2fa4fd2f3408054e62",
15
- "LICENSE_NOTICE.md": "98eb26f7cb4c54956cf442834a1a611ee7d0e148e4f6dee237c0e8f207d7c2e7",
16
- "README.md": "47bd192a8b93cf43dfb2960295c419398d3ff04f85707831d0ce7dc48fb7f006",
17
- "code/language_detection.py": "43dfbf0ad1ef5dd79dac1c37a00b66c739ba84640b26895aec08c9d05fde8c61",
18
  "code/language_detector.json": "6fbf73ef0e200d9fd138e4e53da24bc388239b34b8ba0fc1e40581a108263507",
19
- "code/narrativebench.py": "f01a1e732ab0c91c123555665251d502e3a58940512cd6769f5bd7ea93b799b9",
20
  "code/requirements-language.txt": "6e97b52bbf34a9cfcf579da0c0fc439546e02c96803d4672c3b778955ae9b55a",
21
  "data/documents.parquet": "3224f96d2591957e03a34f0d2447bdf34b71dd99636b4372fea58aaa58103b2e",
22
  "data/evaluation.parquet": "ec10cfdd3ff74c6c10172e71cdd13ac275eff054b2a0eca9e86d0b45803ae352",
 
10
  "files": {
11
  ".gitattributes": "30129f61f2a1b984b358279fb24769ff73836dc7f839a765b67775ae2af04a13",
12
  ".gitignore": "6843cd7b533a96ccb653dd16ece08e028ce16018b537283a1bfa5354c917f028",
13
+ "DATA.md": "4a4fe8900338fcc3d32a5950d513756126e5b69d51eaf9997dd22770adad0057",
14
+ "EVALUATION.md": "abbf35e73452e565b997047081b36a53d82378efead2a00d47d5725c8afc1d63",
15
+ "LICENSE_NOTICE.md": "af69ee567f56f148dfe8a2c344d16deb8b7af812544cd0ac3d89ec8d6fb60a39",
16
+ "README.md": "472c6845c8d107e3f000189f98b1d689204357eeeb022cda603dec8ee813e1f1",
17
+ "code/language_detection.py": "c44fe6841683d4fcb066eec8b4886d7c66e26498086ad59fb3a18f8e6da5ea6b",
18
  "code/language_detector.json": "6fbf73ef0e200d9fd138e4e53da24bc388239b34b8ba0fc1e40581a108263507",
19
+ "code/narrativebench.py": "af470fa92ac17bd3175e35da20c1ec5c3e17b05ae42c3c70d53bd06b0ec31c13",
20
  "code/requirements-language.txt": "6e97b52bbf34a9cfcf579da0c0fc439546e02c96803d4672c3b778955ae9b55a",
21
  "data/documents.parquet": "3224f96d2591957e03a34f0d2447bdf34b71dd99636b4372fea58aaa58103b2e",
22
  "data/evaluation.parquet": "ec10cfdd3ff74c6c10172e71cdd13ac275eff054b2a0eca9e86d0b45803ae352",