nivvis/eq-esconv-sifted

---
license: cc-by-nc-4.0
parent_dataset: thu-coai/esconv
task_categories:
- text-generation
- conversational
tags:
- empathy
- emotional-intelligence
- eq
- emotional-support
- ranked
language:
- en
size_categories:
- 1K<n<10K
---

# EQ-ESConv-Sifted: Elo-Ranked Emotional Support Conversations

The [ESConv](https://huggingface.co/datasets/thu-coai/esconv) dataset (Liu et al., ACL 2021) ranked by empathetic quality via Swiss-style Elo tournament. All 1,300 conversations scored and sorted.

## Why this exists

ESConv is a widely-used emotional support dataset but quality varies significantly — some conversations have excellent empathetic support, others are low-effort or off-topic. This dataset adds Elo rankings so you can filter by quality.

For higher-quality synthetic empathetic conversations, see:
- [nivvis/eq-convos](https://huggingface.co/datasets/nivvis/eq-convos) — 1,774 multi-turn synthetic conversations, Elo-ranked
- [nivvis/eq-dpo](https://huggingface.co/datasets/nivvis/eq-dpo) — 2,880 turn-level DPO preference pairs

This dataset is still useful as:
- Source material for further synthesis
- A benchmark for comparing empathy quality across datasets
- SFT data when filtered to top tiers

## Fields

| Field | Description |
|-------|-------------|
| `id` | Original ESConv index |
| `elo` | Tournament Elo rating (higher = better empathy) |
| `wins` | Tournament wins |
| `losses` | Tournament losses |
| `draws` | Tournament draws |
| `emotion_type` | Primary emotion: anxiety, depression, sadness, anger, fear, etc. |
| `problem_type` | Problem category: breakup, job crisis, ongoing depression, etc. |
| `situation` | Description of the seeker's situation |
| `dialog` | Full conversation as JSON array of `{text, speaker, strategy}` turns |
| `survey_score` | Original crowdworker survey ratings (empathy, relevance, emotion intensity) |

## Quality Tiers

| Tier | Elo | Count |
|------|-----|-------|
| Top 20% | >= 1530 | ~260 |
| Top 50% | >= 1500 | ~650 |
| Bottom 25% | < 1480 | ~325 |

## Tournament Method

- Swiss-style pairing: similar-Elo conversations matched each round
- 10 rounds, ~650 matches per round
- Logit-probe A/B judging: normalized P(A|{A,B}) from single-token output
- Position bias mitigated via randomized A/B presentation order
- Continuous Elo scoring (no binary win/loss, raw probabilities as match scores)

## Citation

Original ESConv dataset:
```bibtex
@inproceedings{liu-etal-2021-towards,
title = "Towards Emotional Support Dialog Systems",
author = "Liu, Siyang and others",
booktitle = "ACL 2021",
year = "2021"
}
```