WildChat-4.8M: Multi-turn AI Dialogue Data

• WildChat-4.8M is a large-scale conversational dataset characterized by millions of real-world multi-turn exchanges between users and GPT models.
• It employs rigorous filtering and detailed annotation to capture user fluency, feedback mechanisms, and linguistic convergence across diverse languages.
• The dataset’s skewed user activity and templated exchanges necessitate critical interpretation when generalizing insights for typical LLM usage.

WildChat-4.8M is a large-scale, publicly available conversational LLM dataset capturing millions of real-world, multi-turn interactions between anonymous users and GPT-based models via a public HuggingFace Spaces application. This corpus enables the study of user-AI interaction patterns, conversational intent, user fluency, feedback mechanisms, and linguistic accommodation at scale. However, its structure, user composition, and associated limitations necessitate caution in generalizing to broader populations or drawing conclusions about typical LLM usage.

1. Construction, Filtering, and Scale

WildChat-4.8M was collected from March 9, 2023 through July 31, 2025 using a public web frontend that provided free GPT model access in exchange for users’ consent to log their conversations. Each record contains the complete text of user and model turns, a hashed IP address (used as a pseudonymous user ID), and geolocation information (country, state, language) (Hicke et al., 27 May 2026). To exclude bursts of non-human or API-style activity, only conversations prior to September 1, 2024 are used. Filtering removes hashed IPs associated with more than three countries, states, or languages, as well as those exceeding 161 total conversations. The resulting dataset consists of approximately 2,522,330 conversations from 1,830,631 distinct users, making it one of the largest publicly accessible corpora of naturalistic LLM dialogues (Hicke et al., 27 May 2026).

In broader scope, the dataset includes about 4.8 million dialogues ("WildChat-4.8M") totaling approximately 20–30 million model and user turns (Blevins, 28 May 2026, Liu et al., 30 Jul 2025). Dialogue lengths average about 4.2–4.8 turns per participant, with conversations ranging across English, Spanish, French, Italian, Portuguese, Russian, Turkish, and Chinese (Blevins, 28 May 2026). Personal identifying information is stripped using Microsoft Presidio and custom regex procedures over all supported languages (Blevins, 28 May 2026). Only dyadic (user ↔ model) exchanges are retained, and all system messages are excluded. Templated exchanges—dominated by API scripts or repetitive pattern use—are a substantial fraction: the top 10 prompt templates account for over 15% of all conversations; overall, 39% of the data is templated (Hicke et al., 27 May 2026).

2. User Composition and Activity Profiles

WildChat-4.8M users are distributed extremely right-skew: over 60% appear only once, fewer than 5% exceed ten conversations, and under 0.1% surpass fifty. The per-user conversation count follows a power law with exponential cutoff:

$$p(k)\propto k^{-\alpha} \cdot e^{-\lambda k},\quad \text{with } \alpha \approx 2.1$$

User "activity levels" are stratified by the number of unique active days:

• Low activity: 1–10 days
• Middle: 11–25 days
• High: 26+ days

However, very few users fall into the high-activity group; the majority engage only transiently (Hicke et al., 27 May 2026). An activity score $A_i = \log(1 + \mathrm{convos}_i)$ is used to normalize for heavy-tailed outliers.

For retention, the daily user return rate $$R(t) = |\mathcal{U}_t \cap \mathcal{U}_{t-1}|/|\mathcal{U}_{t-1}|$$ hovers near 0.10, indicating that only ~10% of a given day's users return the following day. This is considerably lower than populations such as Bing Copilot, where $R(t)$ is approximately 0.30 (Hicke et al., 27 May 2026).

Nearly all user demographics—age, gender, profession—are unavailable; only hashed IDs and geolocated metadata are provided. The dataset is notably skewed toward high-activity ("power") users and repeated/templated usage, with many highly active "users" behaving more like API or automation clients than humans (Hicke et al., 27 May 2026).

3. Annotation Schemes, Data Splits, and Features

WildChat-4.8M has been extended and selectively annotated for research targeting user fluency, task complexity, failures, and feedback (Potts et al., 28 Apr 2026, Liu et al., 30 Jul 2025). The available splits from Zhao et al. are as follows:

Split	Number of Dialogues	Proportion
Train	4,032,000	84%
Valid	384,000	8%
Test	384,000	8%

The “non-toxic” subset (used for fluency/failure studies) includes 3,199,860 dialogues. Manual and automatic annotation protocols have been developed:

• Fluency: Four-tiered levels (minimal, low, moderate, high). Augmentative interaction (collaborative iteration) is almost exclusive to high-fluency users ($\sim$93%), nearly absent among minimal/low-fluency users (Potts et al., 28 Apr 2026).
• Task Complexity: Confidence scores (1–5) plus labels for cognitive complexity, domain expertise, scope ambiguity, and novelty.
• Failure Modes: LLM-driven pipelines tag failures as "visible," "invisible," "mixed," or "none," with invisible failure archetypes including "confidence trap," "silent mismatch," "walkaway," etc. (Potts et al., 28 Apr 2026).
• Implicit Feedback: Each user turn (except the first) is flagged for the presence of feedback, which is further classified (positive, rephrasing, awareness, clarification, correction) (Liu et al., 30 Jul 2025). Manual “dense” annotation of 34 conversations shows turn-wise feedback rates close to unity by the fifth turn.

Average per-conversation metrics include:

• Length: 4.2–6.0 turns per side (σ≈2.1)
• Feedback: Per turn, $P(\text{feedback}\mid\text{turn}=2)=0.88$, rising to $$P(\text{feedback}\mid\text{turn}\ge 5) \approx 0.99$$ (Liu et al., 30 Jul 2025)
• Mean feedback length: 10.3 words (σ≈3.8)

Annotations are produced by LLM pipelines; no standard inter-annotator reliability metrics are available for automated labels (Potts et al., 28 Apr 2026).

4. Patterns of User Fluency, Failure, and Outcomes

WildChat-4.8M enables detailed exploration of the relationship between user proficiency (“AI fluency”), conversational strategy, task complexity, and outcomes (Potts et al., 28 Apr 2026). Key findings:

• High-fluency users attempt more complex tasks (mean complexity rises from ~1.5 for minimal to ~3.1 for high fluency), iterate more, and pursue augmentative engagement rather than “one-shot” delegation. However, overall high fluency is rare: minimal ≈45%, low ≈35%, moderate ≈18%, high ≈2%.
• Failure rates rise with fluency—24% for minimal-fluency, 64% for high-fluency users—but invisible failures (conversations judged “successful” by superficial inspection but actually missing the user goal) are dramatically more frequent for low-fluency users (85.6% of failures are invisible at minimal fluency vs. 22.1% at high fluency).
• High-fluency users are strongly associated with “visible failure” and “partial recovery” patterns; minimal-fluency users most often experience silent “walkaway” failures.
• Augmentative style correlates with high fluency: iterative refinement and proactive engagement dominate for proficient users, while passive “delegation” defines the bulk of low-fluency transcripts.
• The “paradox of AI fluency” is observed: increased engagement yields both more recoverable failures and greater ultimate success, reframing recommendations for practitioner guidance and UI design (Potts et al., 28 Apr 2026).

5. Implicit Feedback and Modeling Applications

WildChat-4.8M’s scale and turn-level feedback flags support large-scale studies of implicit user feedback for model training (Liu et al., 30 Jul 2025). Key insights:

• Feedback turn rate is extremely high: by turn 5, ≥99% of continued dialogues exhibit a feedback event.
• Negative feedback dominates; positive-only feedback is rare (~2%).
• The utility of feedback for SFT is tied to prompt quality (Spearman’s ρ≈0.24, p<0.01). Incorporating semantic content of feedback into SFT models yields measurable performance gains (+0.67 on MTBench for Vic-7B), but degrades performance on more complex, multi-turn tasks (WildBench: –5 points) (Liu et al., 30 Jul 2025).
• Labeling is noisy; auto-detected feedback is accurate at only ~41% (Precision 61.1%, Recall 35.9%).
• Domain coverage is mostly general-purpose or “consumer” tasks; technical queries are sparse.

Best practices for use: pre-filter by prompt quality for fine-tuning, upweight turns 2–3 for feedback, and combine manual with automatic feedback labeling (Liu et al., 30 Jul 2025).

6. Linguistic Accommodation and LLM-Human Symmetry

WildChat-4.8M is central to large-scale studies of linguistic convergence between humans and LLMs (Blevins, 28 May 2026). Principal findings:

• LLMs exhibit pronounced overconvergence toward user style, both in open-class (noun) and LIWC function-word features, across all studied languages. For English, LLMs converge on noun choice ($0.442$) more than three times as frequently as users ($0.141$). For function word categories, LLM→user convergence is twice that of user→LLM (LIWC: $0.068$ vs $A_i = \log(1 + \mathrm{convos}_i)$0).
• Human users accommodate style no more (or less) than they do in human-human corpora. WildChat user convergence matches DailyDialog and Ubuntu Dialogue Corpus baselines.
• Overconvergence is stable across turn positions and language; model rates exceed user rates by 95–330% in all eight tracked languages.
• This asymmetry indicates that LLMs “overfit” to user style, despite humans treating LLMs linguistically much like any conversational partner (Blevins, 28 May 2026).

Feature	WildChat LLM	WildChat User	Δ (LLM–User)
NOUN	0.442	0.141	0.301
LIWC function	0.068	0.035	0.033

7. Biases, Representativeness, and Research Recommendations

WildChat-4.8M, while extensive, is heavily skewed—high-activity and templated users are vastly overrepresented relative to typical model usage (Hicke et al., 27 May 2026). Key caveats and implications:

• Average sentence length and message count per conversation do not increase with user activity, distinct from other datasets (e.g., Bing Copilot), suggesting the high-activity stratum is less human-like.
• User behavior shows “sticky habits”—within-user feature variance matches interday variance, and most users exhibit little temporal evolution in complexity or engagement.
• Downstream artifacts—training sets, benchmarks, taxonomies—may overweight professional, heavily engineered task patterns while underrepresenting casual, one-off interaction modes.
• The dataset’s opt-in, public-facing nature and presence of large non-representative groups (e.g., “Midjourney” and “Blockman” exogenous patterns) limit its generalizability. “Standard” subset filtering removes such patterns, but analogous behavior may recur in other log data (Potts et al., 28 Apr 2026).

Recommendations:

• Use WildChat-4.8M for studies of advanced prompt engineering, multi-step or pipeline tasks, and high-proficiency feedback loops.
• Exercise caution when generalizing to mass-market, “typical” LLM users or using the data to model adoption trajectories or naturalistic long-tail exploration (Hicke et al., 27 May 2026).
• For training or fine-tuning purposes, split workflows by task complexity and feedback type; use feedback semantics selectively to avoid degrading model performance on complex dialogue tasks (Liu et al., 30 Jul 2025).

WildChat-4.8M remains a significant benchmark for real-world AI conversational behavior, provided that researchers account for its distinctive user profile and structural artifacts in both design and interpretation.