Datasets:

Eedi
/

Question-Anchored-Tutoring-Dialogues-2k

@@ -276,7 +276,7 @@ This dataset originates from conversations between Eedi tutors and students to s
 To facilitate **downsampling**, talk moves were applied using the GPT-based classification model in [Moreau-Pernet et al. (2024)](https://dl.acm.org/doi/10.1145/3657604.3664664). The model was fine-tuned on conversation transcripts from small-group math tutoring sessions.
-To evaluate the model’s generalizability to 1:1 chat-based math tutoring sessions, we manually annotated a validation set of 200 tutor utterances sampled through weighted stratified sampling using the orignal codebook from [Moreau-Pernet et al. (2024)](https://dl.acm.org/doi/10.1145/3657604.3664664). The sample distribution was flattened by 0.8 of the original label distribution represented in the full 4k dialogue dataset in order to validate more examples of minority class labels.
 An error analysis was conducted on all mismatched labels. **All** instances where the model predicted a `<Getting Student to Relate>` label involved a common type of Eedi's word problems involving a conversation between two fictional students. The original codebook was not designed with this unique edge-case in mind and as such we posit the `<Getting Student to Relate>` label may not generalize this to 1:1 context. Considering this limitation, we present an additional results excluding GSR labels and find similar acceptable metrics to the original paper.
@@ -320,38 +320,30 @@ We make three important distinctions:
 3. **NER Model release constraints**: Due to the sensitive nature of the labeling task, the fine-tuned NER model used in this process cannot be released publicly due to re-identification concerns. It could potentially be used to detect residual PII in datasets processed using the HIPS anonymization method.
-#### Who are the annotators?
-<!-- This section describes the people or systems who created the annotations. -->
-[More Information Needed]
 #### Personal and Sensitive Information
 <!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->
-[More Information Needed]
 ## Bias, Risks, and Limitations
 <!-- This section is meant to convey both technical and sociotechnical limitations. -->
-[More Information Needed]
-### Recommendations
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
-Users should be made aware of the risks, biases and limitations of the dataset. More information needed for further recommendations.
 ## Citation [optional]
-Publication in Progress. Please reach out to Matthew.Zent@eedi.com for citation information.
-## Dataset Card Authors
 Matthew Zent
 ## Dataset Card Contact
-Matthew.Zent@eedi.com

 To facilitate **downsampling**, talk moves were applied using the GPT-based classification model in [Moreau-Pernet et al. (2024)](https://dl.acm.org/doi/10.1145/3657604.3664664). The model was fine-tuned on conversation transcripts from small-group math tutoring sessions.
+To evaluate the model’s generalizability to 1:1 chat-based math tutoring sessions, the 1st author manually annotated a validation set of 200 tutor utterances sampled through weighted stratified sampling using the orignal codebook from [Moreau-Pernet et al. (2024)](https://dl.acm.org/doi/10.1145/3657604.3664664). The sample distribution was flattened by 0.8 of the original label distribution represented in the full 4k dialogue dataset in order to validate more examples of minority class labels.
 An error analysis was conducted on all mismatched labels. **All** instances where the model predicted a `<Getting Student to Relate>` label involved a common type of Eedi's word problems involving a conversation between two fictional students. The original codebook was not designed with this unique edge-case in mind and as such we posit the `<Getting Student to Relate>` label may not generalize this to 1:1 context. Considering this limitation, we present an additional results excluding GSR labels and find similar acceptable metrics to the original paper.
 3. **NER Model release constraints**: Due to the sensitive nature of the labeling task, the fine-tuned NER model used in this process cannot be released publicly due to re-identification concerns. It could potentially be used to detect residual PII in datasets processed using the HIPS anonymization method.
 #### Personal and Sensitive Information
 <!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->
+This dataset contains real student-tutor dialogues. Extensive efforts were made to mitigate privacy risks, but these risks cannot be alleviated completely. Identifying details were removed or anonymized where detected, but some sensitive content may persist. If you have concerns about any content, please contact the first author at [matthew.zent@eedi.co.uk](mailto:matthew.zent@eedi.co.uk).
 ## Bias, Risks, and Limitations
 <!-- This section is meant to convey both technical and sociotechnical limitations. -->
+We present aggregate demographic and behavioral statistics to help downstream users assess potential biases in the dataset. Here we note highlight high-impact biases and limitations for downstream use:
+1. QATD’s focuses on UK students and intentionally excludes US-based learners.
+2. The data reflects real interactions on the Eedi platform, where tutors sometimes manage multiple students simultaneously, especially during peak hours.
+3. The collection period includes growing public exposure to AI chatbots. As such, there are many instances of students expressing skepticism about whether their tutor was a real person.
+4. Our strong commitment to student privacy limits access to individual-level student details, which may constrain downstream use in student modeling tasks.
 ## Citation [optional]
+Publication in Progress. Please reach out to Matthew.Zent@eedi.co.uk for citation information.
+## Dataset Card Author
 Matthew Zent
 ## Dataset Card Contact
+Matthew.Zent@eedi.co.uk