README.md

metadata

dataset_info:
  features:
    - name: id
      dtype: string
    - name: language
      dtype: string
    - name: source
      dtype: string
    - name: source_id
      dtype: string
    - name: seed_prompt
      dtype: string
    - name: smith_id
      dtype: string
    - name: smith_messages
      list:
        - name: role
          dtype: string
        - name: content
          dtype: string
    - name: teacher_id
      dtype: string
    - name: teacher_messages
      list:
        - name: role
          dtype: string
        - name: content
          dtype: string
    - name: annotations
      list:
        - name: annotator_id
          dtype: string
        - name: session_id
          dtype: string
        - name: annotated_at
          dtype: timestamp[s]
        - name: duration_ms
          dtype: int64
        - name: preference_naturalness
          dtype: string
        - name: preference_cultural_appropriateness
          dtype: string
        - name: preference_instruction_following
          dtype: string
        - name: preference_coherence
          dtype: string
        - name: preference_overall
          dtype: string
        - name: comment
          dtype: string
        - name: instance_feedback
          dtype: string
  splits:
    - name: train
      num_bytes: 11191230
      num_examples: 2295
  download_size: 10393026
  dataset_size: 11191230
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
language:
  - ar
  - ceb
  - cs
  - de
  - es
  - fr
  - id
  - ja
  - sw
  - tl
  - yo
license: apache-2.0
task_categories:
  - text-generation
tags:
  - multilingual
  - human-evaluation
  - preference
  - human-feedback
pretty_name: DataSmith-Preference (Human Evaluation)

Website: ljvmiranda921.github.io/datasmith/

DataSmith-Preference-HumanEval

This dataset contains human preference judgments comparing DataSmith against a teacher baseline that generates without tools. Each comparison shows an evaluator two conversations written from the same English seed prompt, one from each system, and asks which they prefer overall and on four aspects. Evaluators were native or fluent speakers of the target language, and the two systems were shown as A and B in a randomized order that was hidden from them. It covers eleven languages: Arabic, Cebuano, Czech, French, German, Indonesian, Japanese, Spanish, Swahili, Tagalog, and Yoruba.

Dataset Summary

• Languages: Arabic (ar), Cebuano (ceb), Czech (cs), French (fr), German (de), Indonesian (id), Japanese (ja), Spanish (es), Swahili (sw), Tagalog (tl), Yoruba (yo)
• Comparisons: 2,295
• Judgments: 2,317
• Evaluators: 24
• Systems compared: DataSmith against the teacher baseline
• Collected: 7 September 2026 to 25 September 2026
• Median time per comparison: 4.2 minutes

Aspects

Evaluators gave a preference for each aspect below, plus one overall preference.

Aspect	Description
naturalness	Sounds natural and human-like, the way a native or fluent speaker would write, without translationese or unnecessary repetition of phrases or ideas.
cultural_appropriateness	References, names, and scenarios fit the language community and feel contextually appropriate.
instruction_following	The assistant follows the instructions and requests the user makes throughout the conversation.
coherence	The conversation stays focused on the topic or transitions logically, without drifting or losing the thread.

Preference Values

Evaluators chose between A and B, and we resolved those choices back to the system they refer to. Every preference_* field takes one of the following values.

Value	Meaning
strong_smith	Strong preference for the DataSmith conversation
smith	Preference for the DataSmith conversation
teacher	Preference for the teacher conversation
strong_teacher	Strong preference for the teacher conversation
tie_both_good	No preference, both conversations judged good
tie_both_bad	No preference, both conversations judged bad

Language Distribution

Language	Comparisons	Judgments
Arabic (ar)	232	232
Cebuano (ceb)	230	230
Czech (cs)	168	168
French (fr)	214	220
German (de)	236	239
Indonesian (id)	193	193
Japanese (ja)	210	212
Spanish (es)	196	196
Swahili (sw)	159	159
Tagalog (tl)	229	230
Yoruba (yo)	228	238
Total	2,295	2,317

Data Sources

The seed prompts are English instructions sampled from the pool used to build the DataSmith training data.

Source	Comparisons
Nemotron Instruction Following Chat	1,554
Dolci Instruct SFT	741

Dataset Structure

Each row is one comparison and contains the following fields:

Field	Type	Description
id	str	Identifier for the comparison
language	str	ISO 639-1 language code, except Cebuano which uses ceb
source	str	Source dataset the seed prompt came from
source_id	str	Identifier from the source dataset
seed_prompt	str	The English seed instruction both conversations were written from
smith_id	str	Identifier of the DataSmith conversation
smith_messages	list	The DataSmith conversation (role and content)
teacher_id	str	Identifier of the teacher conversation
teacher_messages	list	The teacher conversation (role and content)
annotations	list	One entry per evaluator who judged this comparison

Each entry in annotations contains:

Field	Type	Description
annotator_id	str	Anonymized evaluator identifier
session_id	str	Annotation session the judgment was made in
annotated_at	timestamp	When the judgment was submitted
duration_ms	int	Time spent on the comparison, in milliseconds
preference_naturalness	str	Preference on naturalness
preference_cultural_appropriateness	str	Preference on cultural appropriateness
preference_instruction_following	str	Preference on instruction following
preference_coherence	str	Preference on coherence
preference_overall	str	Overall preference
comment	str	Free-text justification, empty when not given
instance_feedback	str	Free-text feedback on the comparison itself, empty when not given

Usage

from datasets import load_dataset

ds = load_dataset("ljvmiranda921/DataSmith-Preference-HumanEval", split="train")

# Filter by language
arabic_ds = ds.filter(lambda x: x["language"] == "ar")

# Overall preference of the first evaluator on each comparison
print(arabic_ds[0]["annotations"][0]["preference_overall"])

# Win rate for DataSmith, counting strong and plain preferences together
judgments = [a for row in ds for a in row["annotations"]]
wins = sum(1 for a in judgments if a["preference_overall"].endswith("smith"))
print(wins / len(judgments))

# Comparisons with more than one evaluator, for agreement analysis
overlapping = ds.filter(lambda x: len(x["annotations"]) > 1)

# Read what evaluators wrote
for a in judgments:
    if a["comment"]:
        print(a["preference_overall"], a["comment"])

Acknowledgements

We are grateful to the evaluators who participated in our human evaluation for their careful and thoughtful judgments. LJVM and AK acknowledge the support of the UKRI Frontier Grant EP/Y031350/1 (EQUATE). Compute was provided by the Isambard AI National AI Research Resource (AIRR) ST/AIRR/I-A-I/1023 and the Microsoft Research Azure AI Grant. LJVM also thanks the team at Microsoft Research Nairobi and members of the Language Technology Lab for discussions and feedback on this work.