For now, I found a few related resources that might be useful:
I did not find an exact duplicate of the proposed resource, but I found several projects that cover different parts of it: NVC-informed conflict annotation, empathy and counseling schemas, paired rewriting, recipient-side evaluation, conversational sycophancy, and the practical mechanics of running a small community annotation project on the Hub.
The closest starting point I found is PersonaConflicts / Words Like Knives, which provides 5,772 simulated dialogues between familiar social partners, relationship backstories, and a human-annotated subset using NVC-informed communication-breakdown labels. The accompanying EMNLP paper may be useful for comparing schemas and identifying what a new dataset would add beyond synthetic conflict dialogues.
A practical default route might be:
- use existing conflict and empathy data to test a draft annotation schema;
- distinguish directly observable text from inferred feelings or needs;
- retain evidence, uncertainty, multiple interpretations, and unaggregated annotations;
- evaluate extraction, rewriting, recipient perception, de-escalation, and sycophancy separately;
- release the schema, annotation guide, difficult examples, and a small pilot before committing to large-scale annotation.
The smallest useful first release may therefore be a dataset repository containing:
- a draft annotation guide;
- 25–100 independently annotated examples;
- raw annotations as well as any adjudicated version;
- examples of uncertainty and disagreement;
- a small controlled evaluation set;
- a Dataset Card documenting provenance, intended use, and limitations.
One distinction seems particularly useful from the beginning: a feeling or need explicitly stated by the speaker is a different kind of label from a feeling or need inferred by an annotator. For inferred fields, it may help to preserve supporting text, confidence, alternative interpretations, and an insufficient context option rather than requiring one definitive answer.
Closest adjacent resources
NVC-informed conflict resources
PersonaConflicts / Words Like Knives
PersonaConflicts appears to be the closest public resource I found. It includes:
- 5,772 simulated dialogues;
- friends, family members, and romantic partners;
- conflict scenarios and relationship backstories;
- turn-level dialogue data;
- separate raw and aggregated human-annotation files;
- NVC-informed constructive and violent-communication categories.
The paper focuses on how relationship backstory changes human and model perceptions of conversational breakdowns. It is not the same as a full OFNR dataset with inferred Need candidates and paired NVC rewrites, but its released files may be useful for testing an annotation guide, finding difficult examples, preserving relationship context, and establishing a baseline before collecting new conversations.
SpeakSoftly
SpeakSoftly is an NVC-based intervention system for couples’ text conflicts. It separates at least two useful functions:
- detecting aggressive language and suggesting a revision;
- helping users inspect feelings and needs for reflection and perspective-taking.
It also compares intervention modes of different depth. One implication for dataset design is that a full OFNR analysis may not be the best intervention for every turn. A lightweight reminder, a rewrite suggestion, and a deeper reflective guide can be treated as different system behaviors rather than one output format.
Prompt-only NVC baselines
Reducing Conversational Escalation with NVC Constraints operationalizes NVC as lightweight process constraints: discouraging blame, attending to emotional experience, and encouraging clarification before advice.
Can Language Model Moderators Improve the Health of Online Discourse? also includes an NVC-prompted moderator alongside other moderation approaches.
These do not test an NVC training dataset, but they provide concrete prompt-only baselines. A later experiment could compare:
- ordinary prompting;
- generic empathy or active-listening prompting;
- NVC process prompting;
- a model trained on NVC-annotated data;
- NVC training plus NVC prompting.
This could help identify which effects require dedicated data and which can already be produced by an instruction.
Adjacent empathy and counseling resources
EPITOME
EPITOME and its repository provide approximately 10,000 mental-health-support post–response pairs annotated with a theory-grounded empathy framework and supporting rationales.
It is not a Need-annotation dataset, but it is a useful precedent for storing both the final label and the exact text supporting it. A similar rationale field could make inferred feelings and needs more inspectable.
AnnoMI
AnnoMI contains 133 professionally transcribed and expert-annotated demonstrations of high- and low-quality Motivational Interviewing.
Motivational Interviewing and NVC are different frameworks, but AnnoMI is a useful example of:
- domain-expert annotation;
- utterance-level behavioral codes;
- high- and low-quality examples;
- retaining a detailed annotation version;
- documenting collection, transcription, annotation, and post-processing.
Emotional Support Conversation
The Emotional Support Conversation dataset and its ACL paper use a theory-informed strategy taxonomy for emotional-support conversations.
It may be useful as a precedent for annotator tutorials, qualification, strategy-level labels, conversation stages, and post-conversation ratings. Its labels should not be treated as NVC labels, but parts of the collection process may be reusable.
NVC-specific schema ideas
OFNR may be one layer rather than the whole schema
The current CNVC Certification Preparation Packet includes distinctions that may be useful in addition to OFNR:
- observation versus evaluation;
- feeling versus thought or interpretation;
- need versus request or strategy;
- request versus demand;
- stimulus versus cause;
- empathy versus fixing, reassuring, advising, or storytelling;
- empathic sensing versus intellectual guessing;
- idiomatic NVC versus formal or formulaic NVC;
- power-with versus power-over interaction.
These could support a layered schema:
| Layer |
Example question |
| OFNR structure |
Is an Observation, Feeling, Need, or Request present? |
| Evidence |
What text supports that label? |
| Explicitness |
Was it stated or inferred? |
| Function |
Is this a request, demand, repair, advice, or validation? |
| Naturalness |
Is it idiomatic communication or a formal template? |
| Recipient effect |
Did it feel like understanding, pressure, or something else? |
This helps avoid treating any sentence containing four formal components as automatically successful NVC.
The Needs Inventory can be a seed vocabulary
CNVC presents its Feelings and Needs Inventory as a starting resource rather than a closed final ontology.
A dataset could therefore preserve the annotator’s own language and an optional mapping:
need_free_text: "being taken seriously"
mapped_need_candidates:
- respect
- understanding
- to_matter
mapping_confidence: 0.55
mapping_status: approximate
Possible fields include:
| Field |
Purpose |
explicit_need |
Need stated directly by the speaker |
candidate_needs |
One or more inferred hypotheses |
need_free_text |
Annotator’s original wording |
canonical_mapping |
Optional inventory mapping |
evidence_span |
Text supporting the interpretation |
confidence |
Annotator confidence |
alternative_interpretation |
Another plausible reading |
speaker_confirmed |
Whether the speaker later confirmed it |
insufficient_context |
No grounded interpretation available |
This would retain culturally or linguistically specific expressions, permit new terms, and allow separate mappings to NVC and SDT if both are included.
Multi-label and span-level annotation
A single utterance can contain several functions at once: an Observation mixed with evaluation, a Feeling mixed with interpretation, an implicit Need, a Request, blame, defensiveness, a question, or an explanation.
It may therefore help to distinguish:
- span-level labels from whole-utterance labels;
- primary and secondary functions;
- explicit and inferred content;
- single-label tasks from multi-label tasks.
Request versus Demand may require multiple turns
The wording of a request may not reveal whether it functions as a demand. The distinction may become visible after the other person declines or proposes an alternative.
A: Would you be willing to send it by Friday?
B: I cannot do Friday, but I can do Monday.
A1: Monday works; thank you.
A2: Why are you always so difficult?
A3: I asked politely. You need to do this by Friday.
This suggests preserving reply structure, turn order, reactions to refusal, alternative proposals, and pressure after no.
Clarification and repair can be valid outputs
An ambiguous Need does not always have to be resolved by classification. Possible actions include:
- reflect only what was explicitly said;
- offer one tentative interpretation;
- offer several candidate interpretations;
- ask a clarification question;
- ask what kind of help the user wants;
- state that the context is insufficient.
| Strategy |
Example |
| Confident inference |
“It sounds like you need respect.” |
| Tentative inference |
“I wonder whether respect might be part of what matters here.” |
| Multiple candidates |
“Could this be about respect, being heard, or having more choice?” |
| Clarification |
“What felt most important to you in that moment?” |
| Grounded reflection |
“You said this happened repeatedly and that you felt frustrated.” |
Research on conversational grounding and repair, such as Grounding Gaps in Language Model Generations and the survey of clarification-question datasets, may provide useful vocabulary and evaluation ideas.
Annotation workflow and disagreement
Independent annotation before model assistance
If an interface displays an LLM-generated Need or Feeling before a person answers, the model suggestion may shape the human judgment.
Just Put a Human in the Loop? found that LLM suggestions in a complex subjective annotation task could shift the resulting label distribution toward the model’s proposals.
A cautious workflow could be:
Pass 1: independent human annotation
- no model-generated label is shown;
- the annotator records labels, evidence, confidence, and uncertainty;
- free-text interpretation can be entered before opening a taxonomy.
Pass 2: optional assisted review
- model suggestions become visible;
- the annotator may revise the answer;
- the original and revised answers are both retained;
- the annotator records why the answer changed.
independent_annotation:
candidate_needs:
- respect
- understanding
confidence: 0.6
model_suggestion:
candidate_needs:
- belonging
reviewed_annotation:
candidate_needs:
- respect
- belonging
changed: true
change_reason: "The context supports both interpretations."
This permits later analysis of whether assistance corrected mistakes, added alternatives, or merely homogenized labels.
Preserve raw annotations before adjudication
For subjective fields, disagreement may indicate:
- unclear guidelines;
- overlapping categories;
- inadequate context;
- difficult examples;
- several legitimate interpretations;
- instability within an annotator’s own judgment.
Looking Beyond the Majority Vote examines modeling individual human judgments rather than reducing everything to a majority label. Work on diagnosing subjective annotation schemas distinguishes unstable criteria from categories that systematically overlap.
A release could contain:
raw_annotations.jsonl
label_distributions.jsonl
adjudicated_annotations.jsonl
need_distribution:
respect: 0.40
understanding: 0.30
autonomy: 0.20
fairness: 0.10
This supports conventional training, soft-label training, disagreement research, and future taxonomy revision.
Pilot the guide, annotators, and interface together
A pilot can measure:
- median time per item and per field;
- frequency of
insufficient context;
- frequency of multiple candidates;
- item- and label-level disagreement;
- changes after guideline revision;
- effects of visible context;
- effects of free-text-first versus taxonomy-first interfaces.
Difficult cases should be included deliberately:
- Feeling explicit but Need unclear;
- several plausible Needs;
- Request/Demand boundaries;
- irony or understatement;
- culturally indirect language;
- polite but evaluative wording;
- empathic but factually ungrounded replies;
- cases that require previous turns.
Repeating a small hidden subset later for the same annotator can help distinguish stable perspective differences from unstable individual judgments.
Rewriting and controlled comparisons
Evaluate more than whether the output sounds NVC-like
A paired rewrite can improve on one dimension and fail on another.
- Framework transfer — Does the rewrite express the intended NVC-related behavior?
- Meaning preservation — Does it retain the factual claim, boundary, disagreement, request, refusal, and commitment?
- Dialogue function — Does a request remain a request and a refusal remain a refusal?
- Naturalness — Does it fit the relationship rather than sound formulaic?
- Grounding — Does it avoid inventing feelings, motives, or Needs?
- Interactional effect — Does it reduce coercion without becoming evasive or excessively agreeable?
Conditioning on Dialog Acts Improves Empathy Style Transfer is relevant because it treats dialogue-act preservation as part of empathetic rewriting. Work on contextual meaning preservation also suggests that isolated sentence similarity may miss changes visible only in conversation context.
Comparison conditions
For each source message, a pilot could compare:
- the original;
- a human NVC rewrite;
- an LLM rewrite using an NVC prompt;
- a generic empathy or active-listening rewrite;
- a plain politeness rewrite;
- a concise clarification question.
This could separate NVC-specific structure from generic warmth, politeness, verbosity, clarification, and dedicated training effects.
Pairwise evaluation
For each criterion:
A is better
B is better
Approximately equal
Both are unacceptable
Cannot determine
Separate questions can ask which candidate:
- preserves meaning better;
- is more natural;
- makes fewer unsupported assumptions;
- contains the clearer, refusal-compatible request;
- better fits the current stage of the conversation.
Candidate order and generation method should be hidden where possible.
Controlled contrast set
A small contrast set may be useful even before a training corpus exists.
| Fixed |
Changed |
| Meaning |
NVC-like versus ordinary wording |
| Formal OFNR structure |
Supportive versus coercive function |
| Warm tone |
Correct versus incorrect factual position |
| Need candidate |
Evidence present versus absent |
| Interpretation |
Tentative versus assertive wording |
| Ambiguous input |
Inference versus clarification |
| Initial request |
Reaction after refusal |
| Content |
Formal versus idiomatic NVC |
Such a set could support annotation calibration, prompt comparison, regression testing, and checking automated judges.
Evaluation map and controls
Avoid one global “NVC quality” score
The following targets are related but not interchangeable:
| Target |
Example question |
| OFNR extraction |
Were the relevant spans identified? |
| Grounded inference |
Is an inferred Need supported by the dialogue? |
| Uncertainty |
Does confidence reflect ambiguity? |
| Rewrite fidelity |
Was the original meaning preserved? |
| Dialogue function |
Did a request, refusal, or question retain its function? |
| Recipient perception |
Did the person feel understood rather than analyzed? |
| Applicability |
Was an NVC-style intervention useful in this turn? |
| Conversation outcome |
Did escalation, clarity, or willingness to continue change? |
| Factual stability |
Did the model retain warranted disagreement? |
| Sycophancy |
Did support become unsupported agreement or reinforcement? |
A high score on one does not imply a high score on the others.
Applicability and support mode
Empathy Applicability Modeling treats the need for an empathic response as a question preceding generation.
Although that work concerns health queries, the distinction may be useful here. A turn may call primarily for:
- information;
- practical coaching;
- emotional relating;
- reflective listening;
- clarification;
- direct action;
- no NVC intervention.
This can help avoid teaching a model to produce a full emotional interpretation when the user wants a concise factual answer.
Expert coding versus recipient perception
Expert NVC coding and recipient experience may be separate targets.
Empathic Conversations separates self-report, counterpart report, and third-party assessment.
| Evaluator |
Best positioned to judge |
| NVC practitioner |
Framework distinctions |
| Original speaker |
Whether an inferred Feeling or Need fits |
| Recipient |
Whether the response feels respectful and understandable |
| Independent annotator |
Observable text and general naturalness |
| Domain specialist |
High-stakes factual or safety content |
| Model judge |
Exploratory scoring, after validation |
Conversation-level outcomes
Turn quality and conversation outcomes should remain separate.
Possible measures include:
- escalation or de-escalation;
- willingness to continue;
- mutual understanding;
- clarity of the disagreement;
- clarity of the next action;
- whether a misunderstanding was repaired;
- agreement, satisfaction, and perceived fairness.
CaSiNo is an adjacent example because it separates objective negotiation outcomes from satisfaction and counterpart perceptions.
These outcomes are affected by many factors besides communication style, so they are evidence about interaction results rather than a direct causal measure of NVC alone.
Sycophancy controls
It may help to separate factual or opinion-following sycophancy from conversational support.
SycophancyEval includes cases where a model is pressured to change an answer.
BenSyc provides a different reference point. In addition to a binary label, its five-class configuration distinguishes:
- Invalidation;
- Neutral;
- Support;
- Validation;
- Escalation.
BenSyc is based on Bengali social contexts, so its taxonomy should not be assumed to transfer unchanged. The reusable idea is the separation: emotional support, strong validation, escalation, and factual agreement need not be one label.
Possible checks include:
- Does the model agree with a false claim?
- Does it change a correct answer after social pressure?
- Does it validate an emotion without endorsing the user’s interpretation?
- Does it intensify hostility toward an absent third party?
- Can it disagree while remaining respectful?
- Can it acknowledge uncertainty rather than align with whichever party is speaking?
LLM-as-judge controls
If an LLM is used for exploratory scoring:
- publish the rubric;
- randomize candidate order;
- conceal which system produced each response;
- test sensitivity to response length;
- repeat a subset across runs;
- compare more than one judge model where practical;
- retain human evaluation for central claims.
Longer or more explicitly therapeutic responses may receive higher automatic empathy scores even when they are less natural or less grounded.
Possible pilot routes
Default: annotation-schema pilot
- Select 25–100 examples from licensed synthetic, role-play, or existing public sources.
- Create a small annotation guide.
- Collect multiple independent annotations.
- Preserve raw labels, evidence, confidence, and disagreements.
- Revise the guide using difficult cases.
- Repeat on a fresh sample.
- Publish the schema, examples, and reliability results.
This route answers:
- Are the fields understandable?
- Which fields are observable or interpretive?
- Which categories overlap?
- How often is context insufficient?
- How long does annotation take?
- Which disagreements can a better guide resolve?
Alternative A: rewrite pilot
Compare:
- human NVC rewrite;
- NVC prompt;
- generic empathy prompt;
- plain politeness prompt;
- clarification-only prompt.
Evaluate meaning, dialogue function, grounding, naturalness, and recipient preference separately.
Alternative B: evaluation-first pilot
Before producing training data, compare current models under ordinary, active-listening, and NVC instructions on:
- PersonaConflicts examples;
- controlled Request/Demand cases;
- ambiguous Need cases;
- clarification cases;
- selected SycophancyEval items;
- a small Support/Validation/Escalation set.
This could identify which failure cases actually require new training data.
Alternative C: documentation-first release
A public artifact could initially contain:
- draft schema;
- annotation guide;
- 10–20 worked examples;
- unresolved edge cases;
- intended uses and non-goals;
- open Community discussions.
That may be enough to obtain useful feedback before collection begins.
Alternative D: controlled contrast set
Create a small evaluation set isolating:
- Observation versus evaluation;
- Feeling versus thought;
- Need versus strategy;
- Request versus Demand;
- empathy versus fixing;
- tentative sensing versus confident guessing;
- NVC form versus interactional function;
- validation versus unsupported agreement.
Hub implementation, governance, and reproducibility
Dataset repository
Hugging Face’s dataset documentation describes Hub datasets as versioned repositories, while the upload guide covers common file-based workflows.
A possible layout:
README.md
CHANGELOG.md
annotation_guide.md
schema/
v0.1.json
v0.2.json
data/
pilot_round_1.jsonl
pilot_round_2.jsonl
raw_annotations.jsonl
label_distributions.jsonl
adjudicated_annotations.jsonl
eval/
extraction.jsonl
rewrites.jsonl
controlled_contrasts.jsonl
sycophancy_controls.jsonl
docs/
difficult_examples.md
decisions.md
collection_and_consent.md
removal_policy.md
Dataset Card
The Dataset Cards documentation recommends documenting content, context, responsible use, bias, license, language, and size.
Additional sections could include:
## NVC framework and references
## Label epistemology
- directly observable labels
- annotator inferences
- speaker-confirmed labels
- suggested rewrites
## Data provenance
- real conversation
- consented elicitation
- role-play
- expert-written vignette
- synthetic dialogue
- human or model rewrite
## Context boundaries
- previous turns shown
- relationship information shown
- whether both sides are represented
## Annotation procedure
- annotator background
- independent and assisted passes
- adjudication
- raw-label availability
## Intended use
## Out-of-scope use
- clinical diagnosis
- determining a person’s “true” needs
- autonomous high-stakes mediation
- legal or employment decisions
## Corrections and removals
## Version and migration history
Datasheets for Datasets is another useful framework for motivation, composition, collection, uses, and maintenance.
Community discussions
Hub repositories support Discussions and Pull Requests.
Possible discussions:
[DRAFT] Explicit versus inferred Needs
[DRAFT] Multiple Need candidates and confidence
[OPEN] NVC-to-SDT mappings
[OPEN] Meaning-preservation rubric
[OPEN] Request versus Demand examples
[POLICY] Real-dialogue consent and removal
[RELEASE] Pilot schema v0.2
This allows contributors to focus on individual design questions rather than agree on the entire project at once.
Annotation interfaces
Argilla is one Hub-connected option for labels, ratings, rankings, and free text.
An interface could include:
- conversation context;
- target utterance;
- span labels;
- multi-label Feeling and Need candidates;
- confidence;
- rationale;
- alternative interpretation;
- rewrite;
- pairwise rewrite ranking.
The interface itself is part of the protocol. Label order, visible context, preselected candidates, and model suggestions can affect the collected labels.
Lower-risk starting sources
Initial schema development could use:
- appropriately licensed synthetic corpora;
- expert-written fictional examples;
- controlled role-play;
- purpose-built contrast cases.
This avoids beginning with private family, partner, workplace, or counseling conversations before the protocol and governance process are stable.
Provenance and licensing
Useful metadata include:
record_id: nvc_000001
conversation_id: conv_00014
source_type: synthetic_conflict
source_record_id: original_dataset_123
original_language: en
schema_version: 0.2
guideline_version: 0.3
annotation_round: pilot_2
source_license: MIT
redistribution_reviewed: true
If an existing corpus is annotated or rewritten, it is useful to check:
- whether its text can be redistributed;
- whether annotations alone can be released;
- whether derivative rewrites can be released;
- whether training and commercial use are permitted.
The Hub provides an overview of repository licenses, but each source’s own terms still need to be checked.
Human-participant collection
If new conversations are collected, useful questions include:
Before collection
- What uses are covered: public display, training, evaluation, demonstration?
- Who may access the data?
- How are third-party statements handled?
- What are the withdrawal conditions?
- Is ethical or institutional review appropriate?
During collection
- Can participants skip questions or stop?
- How is sensitive content handled?
- Is identifying information minimized and separated?
- How are annotators protected from difficult material?
After collection
- What is publicly released?
- How are corrections and removals handled?
- What are the limits of retracting downloaded copies?
- What are the limits of removing influence from trained models?
Human Participants in AI Research provides a broader guide to consent, autonomy, compensation, harm reduction, and reporting.
Gated datasets can help manage access, but gating does not replace consent, rights review, anonymization, or a removal process.
Versioning and split hygiene
Useful version fields include:
dataset_version: 0.3.0
schema_version: 0.3
guideline_version: 0.4
annotation_round: pilot_2
Train, development, and test splits should be grouped by conversation and source family rather than randomly by utterance.
Possible grouping keys:
conversation_id:
scenario_id:
source_record_id:
speaker_pair_id:
rewrite_family_id:
synthetic_template_id:
This reduces leakage from adjacent turns, several rewrites of one message, paraphrased scenarios, repeated speakers, or shared synthetic templates.
If the resource later becomes a benchmark, it may also help to separate a public development set from a frozen or periodically refreshed evaluation set.
In short, many of the pieces needed to test the idea already exist without first building the full corpus. PersonaConflicts provides a close conflict-data starting point; EPITOME, AnnoMI, and ESConv provide annotation-process precedents; CNVC materials provide distinctions beyond OFNR; SycophancyEval and BenSyc provide different controls for agreement and support; and the Hub provides repository, documentation, discussion, versioning, and annotation tooling.
A particularly reusable first contribution might be a small public pilot preserving:
- explicit versus inferred content;
- evidence and uncertainty;
- multiple Need candidates;
- raw annotator disagreement;
- clarification as an alternative to guessing;
- meaning preservation in rewrites;
- separate evaluation of support, recipient perception, de-escalation, and factual stability.
That would make it easier to estimate annotation cost, revise the schema, identify where dedicated training data is actually useful, and give future collaborators something concrete to inspect or extend.