pretty_name: "Cognitive Reframing Pairs: a synthetic dataset with contrasts across seven linguistic attributes"
license: cc-by-4.0
language:
- en
task_categories: - text-generation
size_categories: - 1K<n<10K
tags: - cognitive-reframing
- cognitive-behavioral-therapy
- psychotherapy
- psychology
- mental-health
- synthetic
- activation-steering
configs: - config_name: default
data_files:- split: train
path: pairs.jsonl
- split: train
Cognitive Reframing Pairs: a synthetic dataset with contrasts across seven linguistic attributes
Cognitive Reframing Pairs is a synthetic English corpus for cognitive reframing. It contains 605 situations and 1,096 original thoughts, with 2,000 contrastive pairs of reframes. Each pair targets a difference in one linguistic attribute, such as empathy or actionability.
Cognitive reframing is a technique used in cognitive behavioral therapy to reconsider how we interpret a situation. Given a situation and a thought, the task is to generate an alternative thought that respects the facts. It may question an assumption, acknowledge an emotion, or suggest a response.
The dataset's attributes and thought categories were inspired by Sharma et al. (ACL 2023). All examples were newly generated.
Try thought steering, a browser demo that uses this dataset for activation steering: adjust how much Qwen3.5-0.8B reframes negative thoughts and compare responses with steering off.
Example
A pair with different levels of empathy:
| Field | Text |
|---|---|
| Situation | I laid out my paints and brushes on the kitchen table to finish the watercolor I want to enter in the community art fair. The show is next Saturday, and I've been working on the same picture every night this week. |
| Thought | I've spent every night this week on this one painting, but it still isn't done. Maybe I'm not even a real artist at all - a worthless fraud, nothing more. |
| Less empathic | I've put most of this week into this one painting, and it still isn't done. But one painting that isn't finished doesn't mean I'm not a real artist - it just means I haven't completed this particular piece. |
| More empathic | I've put most of this week into this one painting, and it still isn't done, and that really stings. But one painting that isn't finished doesn't mean I'm not a real artist - it just means I haven't completed this piece, and it makes sense to feel frustrated about it. |
Both less and more are reframes.
Dataset structure
pairs.jsonl contains 2,000 rows, one self-contained contrastive pair per row. All fields are strings.
| Field | Content |
|---|---|
pair_id |
Unique pair identifier |
situation_id, thought_id |
Shared situation and thought identifiers |
situation |
Situation text |
thought |
Original thought |
domain |
Life domain |
condition |
Thought-generation category |
attribute |
Reframing attribute |
less, more |
Two reframes, giving 4,000 reframes in total |
persona |
Exact persona context used to generate the situation |
persona_source_id |
Source persona identifier in Nemotron-Personas-USA |
Situation and thought text repeat when pairs share an input. Their identifiers preserve these relationships.
The situations span nine life domains:
- Work and education
- Family
- Romantic relationships
- Friendships and social life
- Finances and housing
- Health and body
- Daily responsibilities
- Leisure and personal goals
- Identity and belonging
Thought generation uses the 13 categories listed in Appendix D of Sharma et al. (2023): all-or-nothing thinking, overgeneralizing, labeling, fortune telling, mind reading, emotional reasoning, should statements, personalizing, disqualifying the positive, catastrophizing, comparing and despairing, blaming, and negative feeling or emotion. The last category allows distress without assuming a cognitive distortion.
The dataset covers seven linguistic attributes (285–286 pairs each):
| Attribute | Meaning of more |
|---|---|
thinking_trap_addressing |
More directly questions the thought's biased inference or rigid rule. |
rationality |
Gives stronger reasons for the alternative interpretation. |
positivity |
Emphasizes plausible favorable possibilities. |
empathy |
Expresses more understanding or self-compassion. |
actionability |
Makes a feasible response clearer. |
specificity |
Refers more closely to the situation and thought. |
readability |
Uses easier wording and sentence structure. |
Data generation
I generated the corpus with Ornith-1.5-9B (Q5_K_M) in llama.cpp, with reasoning enabled and JSON output.
- Select 500 adult personas from NVIDIA's Nemotron-Personas-USA to provide context.
- Generate a situation for a life domain, then up to two thoughts under different categories.
- Generate each reframe pair jointly for a target attribute, with up to two attributes per thought.
Sampling favors categories with fewer accepted examples. Automatic filters check output format, sentence count, exact duplicates, and lexical overlap with the original thought. The released corpus contains examples derived from 491 personas.
The generation repository provides the code, prompts, configuration, and reproduction instructions.
License
CC BY 4.0. Created by Gustave Cortal, using persona contexts derived from NVIDIA's Nemotron-Personas-USA, also under CC BY 4.0. The generation code is MIT-licensed.