README.md

---
language:
- en
license: apache-2.0
task_categories:
- conversational
- text-generation
tags:
- reinforcement-learning-from-human-feedback
- dialogue
- conversational-ai
- preference-alignment
dataset_info:
- config_name: feedback_decoder_binary
  features:
  - name: predictions
    dtype: int64
  - name: labels
    dtype: int64
  - name: game_turn_id
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 5115528
    num_examples: 7003
  download_size: 651037
  dataset_size: 5115528
- config_name: feedback_decoder_ternary
  features:
  - name: predictions
    dtype: int64
  - name: labels
    dtype: int64
  - name: game_turn_id
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 4936671
    num_examples: 7003
  download_size: 640213
  dataset_size: 4936671
- config_name: human_eval
  features:
  - name: correctness
    dtype: string
  - name: feedback
    dtype: string
  - name: delta_clicks
    sequence: string
  - name: game_id
    dtype: string
  - name: turn_id
    dtype: int64
  - name: round
    dtype: int64
  - name: policy_name
    dtype: string
  - name: targets
    sequence: string
  - name: selected
    sequence: string
  - name: deselected
    sequence: string
  - name: context
    sequence: string
  - name: end
    dtype: int8
  - name: currently_selected
    sequence: string
  splits:
  - name: train
    num_bytes: 3190406
    num_examples: 8111
  download_size: 347537
  dataset_size: 3190406
- config_name: interaction
  features:
  - name: game_id
    dtype: string
  - name: turn
    dtype: int64
  - name: end
    dtype: string
  - name: s comments
    dtype: string
  - name: speaker
    dtype: string
  - name: len
    dtype: string
  - name: clicks
    sequence:
      sequence: string
  - name: context
    sequence: string
  - name: targets
    sequence: string
  - name: chat
    sequence: string
  - name: dataset_alias
    dtype: string
  - name: policy_name
    dtype: string
  - name: date
    dtype: date32
  - name: round
    dtype: int64
  splits:
  - name: train
    num_bytes: 8775426
    num_examples: 7920
  download_size: 2120499
  dataset_size: 8775426
- config_name: turn
  features:
  - name: chats
    sequence: string
  - name: clicks
    sequence:
      sequence: string
  - name: targets
    sequence: string
  - name: game_id
    dtype: string
  - name: end
    dtype: int8
  - name: context
    sequence: string
  - name: turn_id
    dtype: int8
  - name: currently_selected
    sequence: string
  - name: deselected
    sequence: string
  - name: selected
    sequence: string
  - name: chat_feedback
    dtype: string
  - name: game_turn_id
    dtype: string
  - name: prob_action
    dtype: float64
  - name: dataset_alias
    dtype: string
  - name: policy_name
    dtype: string
  - name: date
    dtype: date32
  - name: round
    dtype: int64
  splits:
  - name: train
    num_bytes: 45674609
    num_examples: 59431
  download_size: 6438021
  dataset_size: 45674609
configs:
- config_name: feedback_decoder_binary
  data_files:
  - split: train
    path: feedback_decoder_binary/train-*
- config_name: feedback_decoder_ternary
  data_files:
  - split: train
    path: feedback_decoder_ternary/train-*
- config_name: human_eval
  data_files:
  - split: train
    path: human_eval/train-*
- config_name: interaction
  data_files:
  - split: train
    path: interaction/train-*
- config_name: turn
  data_files:
  - split: train
    path: turn/train-*
---

# Retrospective Learning from Interactions (Respect) Dataset

This dataset supports **Retrospective Learning from Interactions (Respect)**, a paradigm introduced in the paper [The Era of Real-World Human Interaction: RL from User Conversations](https://huggingface.co/papers/2509.25137).

The paper introduces Reinforcement Learning from Human Interaction (RLHI), a novel approach that learns directly from in-the-wild user conversations. This enables continual model improvement and multifaceted alignment of conversational models, moving beyond traditional pre-annotated, expert-generated human feedback. The dataset facilitates two complementary methods: RLHI with User-Guided Rewrites and RLHI with User-Based Rewards, linking long-term user personas to turn-level preferences.

*   **Paper**: [The Era of Real-World Human Interaction: RL from User Conversations](https://huggingface.co/papers/2509.25137)
*   **Project Page**: [https://lil-lab.github.io/respect](https://lil-lab.github.io/respect)
*   **GitHub Repository**: [https://github.com/lil-lab/respect](https://github.com/lil-lab/respect)

## Sample Usage

You can load the data and associated checkpoints as follows:

```python
from datasets import load_dataset
from transformers import Idefics2ForConditionalGeneration
from peft import PeftModel
import torch # Ensure torch is imported

# Download data
ds = load_dataset("lil-lab/respect", name="turn", split="train")

# Download checkpoints
checkpoint = "HuggingFaceM4/idefics2-8b"
model_id = 'lil-lab/respect'

model = Idefics2ForConditionalGeneration.from_pretrained(
    checkpoint, torch_dtype=torch.bfloat16)
peft_model = PeftModel.from_pretrained(
    model, model_id, adapter_name="r6_bp", revision="r6_bp")
```