alignment-papers-text/2304.07327_OpenAssistant_Conversations_--_Democratizing_Large.md

### **OpenAssistant Conversations - Democratizing Large** **Language Model Alignment**

**Andreas Köpf** _[∗]_ **Yannic Kilcher** _[∗]_
```
     andreas.koepf@provisio.com yannic@ykilcher.com

```

**Dimitri von Rütte** **Sotiris Anagnostidis** **Zhi-Rui Tam** **Keith Stevens**


**Abdullah Barhoum** **Nguyen Minh Duc** **Oliver Stanley** **Richárd Nagyfi** **Shahul ES**


**Sameer Suri** **David Glushkov** **Arnav Dantuluri** **Andrew Maguire**


**Christoph Schuhmann** **Huu Nguyen**


**Alexander Mattick**
```
           alexander.mattick@googlemail.com

```

**Abstract**


Aligning large language models (LLMs) with human preferences has proven
to drastically improve usability and has driven rapid adoption as demonstrated
by ChatGPT. Alignment techniques such as supervised fine-tuning ( _SFT_ ) and
reinforcement learning from human feedback ( _RLHF_ ) greatly reduce the required
skill and domain knowledge to effectively harness the capabilities of LLMs,
increasing their accessibility and utility across various domains. However,
state-of-the-art alignment techniques like _RLHF_ rely on high-quality human
feedback data, which is expensive to create and often remains proprietary. In an
effort to democratize research on large-scale alignment, we release OpenAssistant
Conversations, a human-generated, human-annotated assistant-style conversation
corpus consisting of 161,443 messages in 35 different languages, annotated with
461,292 quality ratings, resulting in over 10,000 complete and fully annotated
conversation trees. The corpus is a product of a worldwide crowd-sourcing effort
involving over 13,500 volunteers. Models trained on OpenAssistant Conversations
show consistent improvements on standard benchmarks over respective base
models. We release our code [2] and data [3] under a fully permissive licence.


A list of contributors who have chosen to be acknowledged by name can
be found at `[https://open-assistant.io/contributors](https://open-assistant.io/contributors)` .


_∗_ These authors contributed equally to this work.
2 `[https://github.com/LAION-AI/Open-Assistant](https://github.com/LAION-AI/Open-Assistant)`
3 `[https://huggingface.co/OpenAssistant/oasst1](https://huggingface.co/OpenAssistant/oasst1)`


37th Conference on Neural Information Processing Systems (NeurIPS 2023) Track on Datasets and Benchmarks.


**1** **Introduction**


Artificial intelligence (AI), particularly in the field of natural language processing, has witnessed
rapid progress in recent years. Major advancements are primarily driven by a straightforward formula:
take a Transformer [1]-based architecture, increase the parameter count by enlarging depth and width,
increase the size of the training corpus, and increase the scale of training compute. Although models
have for some time exhibited an extraordinary ability to fit the training data and generalize based on
their trained objective [2, 3], their adoption among the general public has until recently been slow.
This can be mainly attributed to misalignment between model predictions and final intended usage.


The alignment of AI systems to human values, intentions, and preferences is a vital and intricate
challenge within the AI research domain. This refers to the process of ensuring that AI systems
can not only successfully optimize the provided surrogate training objectives, but also that their
predictions are in line with their intended purpose and adhere to ethical and safety standards provided
by humans [4, 5]. One possible solution is assistant-style fine-tuning of language models that has
recently emerged as a promising approach to making large language models more in line with human
preferences by generating more desirable outputs based on explicitly collected human preference
data [6, 7, 8, 9, 10, 11] and thus making them more useful.


A notable instance of such an assistant-style model is ChatGPT, which has gained unprecedented
user growth due to remarkable capabilities demonstrated in a wide range of fields, but also easeof-use for the end user [12]. Aligning the model’s predictions is in this case accomplished by
introducing human-generated examples of intended usage and using reinforcement learning from
human feedback [13, 14]. In _RLHF_, the human acts as a teacher and provides feedback in the form of
rewards or penalties. In more detail, Ouyang et al. [13] proposed a three stage procedure to align
language models: First, collect human-generated demonstrations of desired behaviour and train a
supervised fine-tuned ( _SFT_ ) model. Second, train a reward model (RM) on human-annotated rankings
for different model outputs. Third, use the RM as a reward function and fine-tune the _SFT_ model to
maximize the reward generated by its responses. This is achieved using the PPO algorithm [15].


It becomes apparent that the benefits of all the aforementioned stages are predominantly dependent
on the quality of the data used [16]. Despite this, availability of large-scale human feedback datasets
for the open research community remains scarce. Most openly accessible datasets are comprised of
synthetic data of instructions automatically generated by querying language models [17, 18, 19, 20].
Unfortunately, these datasets are limited with respect to their complexity, creativity and quality, as
they rely on a pre-specified list of possible instruction types. Other datasets, such as Vicuna [21], use
human-generated instructions, but still rely on langauge models to produce the respective responses.
Without sufficiently broad and high quality data, even models with substantial size and pre-training
would be inadequate for building capable, helpful, and harmless AI assistants.


Research in this area has predominantly been confined to a select few research labs with access to the
required resources to engage in large-scale training and data collection. This monopolization of access
to quality data undermines the potential for inclusive and diverse research endeavours, particularly in
relation to alignment challenges, which arguably constitute some of the most crucial research areas
of our time. In an effort to democratize research on aligning large language models, we introduce
and release the OpenAssistant Conversations dataset. This dataset is the culmination of an extensive
open- and crowd-sourcing initiative, and its release to the research community seeks to promote
more inclusive research in this highly-influential domain. We provide a comprehensive analysis of
the dataset, assessing ethical implications and safety considerations. We also fine-tune and release
several assistant and preference models to further advance open access and research in this area. This
transparency allows for iterative improvements on the released artifacts, fostering a more collaborative
and inclusive research environment. By providing such a large and diverse dataset, OpenAssistant
Conversations opens up new avenues of research in the field, enabling researchers to explore the
complexities of human language and interactions in ways that were not possible before [22]. In the
following sections, we delve into the intricacies of the OpenAssistant Conversations dataset and
discuss its implications for the alignment of large language models and for society at large.


2


**2** **Data Format**


The proposed dataset consists of a list of conversations. The basic data structure is a _Conversation_
_Tree (CT)_, with nodes representing written messages in a conversation. A CT’s root node represents
an initial prompt, given by the prompter. To avoid confusion, we call the roles of the conversation
_prompter_ and _assistant_ . This allows us to reserve the term _user_ for the human contributors. Although
our efforts focus largely on human contributions, both the prompter and assistant roles can, in
principle, be fulfilled by either a human user or a machine. Every tree node is labelled by its role,
and can have multiple children of the opposite role, each of which represents a separate next step in
the conversation. A path from the root to any node in the tree (including to itself) is called a _thread_,
and it represents a valid conversation with the prompter and the assistant taking turns. Tree nodes
are annotated with additional data such as user-provided labels and metadata, such as collection
timestamp and indicated language. Each _assistant_ node further has a rank associated which orders it
compared to replies of the parent prompt, according to user preferences.


Figure 1: An example CT of depth 4 containing 12 messages. Any path from the root prompt to a
node is a valid thread.


**3** **Data Collection**


The OpenAssistant Conversations dataset is a comprehensive collection of conversational data that
was obtained through a crowd-sourcing effort involving more than 13,000 volunteers. The data
was collected using a web-app interface, dividing the collection of each tree into five separate
steps: _prompting_, _labelling prompts_, _adding reply messages as prompter or assistant_, _labelling replies_,
and _ranking assistant replies_ . Users were fully informed about contributing to a public dataset. The
dataset was curated with content moderation and spam filtering as key components of the annotation
pipeline, ensuring high quality and safety standards.


Volunteers completed over 625,000 tasks in total, resulting in the collection of over 10,000 fully
annotated and filtered Conversation Trees. Example User Interface (UI) displays of the data collection
platform can be found in Appendix C, and current collection parameter settings, such as the number
of collected replies to any parent message, can be found in Appendix G. In the following sections, we
provide more details regarding the various aspects of the data collection pipeline.


**3.1** **Single-Step Collection**


Data collection is structured to be both efficient and effective by breaking the work down into single
units and advancing multiple conversation trees one step at a time. This approach minimizes data
loss due to user attrition and ensures that every unit of work is captured for utilization. The users
are presented with a range of task types, either by choice or through random sampling (weighted
according to current requirements). The task types include creating prompts, replying as an assistant,
replying as a prompter, labeling prompts or replies, and ranking prompter or assistant replies.


**Create a prompt.** Users write an initial prompt that forms the root of a new conversation tree. As
this task is particularly popular, a lottery system manages the selection of new prompts, with only a
fixed number of prompts being chosen for continuation at any given moment. This method serves to
regulate the influx of new prompts and maintain a balanced distribution of tasks.


**Reply as assistant.** Replying as an assistant is a more labor-intensive task that necessitates users
to carefully consider their responses and often engage in external research to provide a helpful and


3


relevant answer to the prompter’s request. This task type, despite its demanding nature, has been
reported to be the most enjoyable by many users due to the diverse array of topics covered. To account
for the increased effort required for this task, a reward system has been implemented to incentivize
users to participate. See Figure 6 for a UI preview.


**Reply as prompter.** The task of replying as a prompter, on the other hand, does not impose strict
quality requirements but instead emphasizes on the importance of diversity to accommodate various
use-cases. Examples of prompter replies may include asking for clarification, modifying the original
intent, posing a follow-up question, or changing the direction of the conversation altogether.


**Label a prompt or reply.** Users are presented with a message from the database along with the
preceding conversation thread (if available) and are asked to categorize the message according to
three dimensions: spam detection, guideline adherence, and quality. For spam detection, users assess
whether the message is unsuitable for inclusion in the dataset, for instances of obvious spam or
trolling. Messages flagged as spam by multiple users are automatically removed from the dataset.


Guideline adherence is evaluated through a set of labels that determines whether the contribution
aligns with the established guidelines (see Figure 4). These labels encompass the message being in a
language other than the specified one, containing personally identifiable information, hate speech,
sexual content, or being deemed inappropriate. Messages labelled in this manner are subsequently
reviewed by human moderators.


Quality labels require users to rate the message on a five-point _Likert_ scale across dimensions such
as quality, creativity, humorousness, politeness, and harmlessness. These labels are stored for later
analysis and application. Notably, users can voluntarily assign these labels (as well as spam &
guideline adherence labels) to any message within the system, even as part of another task, as an
additional contribution.


**Rank assistant replies.** Users are presented with two or more responses to the same parent message
and asked to rank them in order of preference. This allows for a comparative analysis of the various
responses and helps in identifying the most effective and engaging replies (Figure 5).


In summary, this data collection methodology effectively divides work into single units, minimizes
data loss due to user attrition, and captures valuable information for future analysis and application.
By offering users a diverse range of task types, the study encourages active participation and ensures
the collection of rich and varied data for a comprehensive understanding of the subject.


**3.2** **Message Tree State Machine**


The tree state machine serves as a systematic approach to managing the progression of message trees
throughout the data collection process. This method ensures that each tree undergoes a series of
states until it reaches completion, beginning with the creation of new trees by randomly sampling
from the pool of initial prompts. The various states that a message tree passes through include the
_initial prompt review state_, _growing state_, and _end state_, as well as the _aborted low-grade state_ for
trees that are deemed unsuitable for inclusion in the dataset and the _halted by moderator_ state for
trees that have manually been halted by a community moderator.


Upon the creation of a new tree, it enters the _initial prompt review state_, where multiple users are
tasked with providing labels to assess its quality and suitability. This state plays a crucial role in
identifying any potential issues with the initial prompt, and demands special attention, as the entire
rest of the tree (potentially several dozens of tasks) is rooted in the initial prompt. If, at this point,
the provided labels indicate that the tree contains spam or unsuitable content, it is transitioned to the
_aborted low-grade state_ and subsequently removed from the dataset. Conversely, if the tree passes
the _initial prompt review state_, it proceeds to the _growing state_ .


The _growing state_ involves the continuous issuance of tasks to users, such as providing replies, labels,
and rankings, to facilitate the development and expansion of the conversation tree. This state is
essential for collecting diverse and rich data, as it allows for the accumulation of multiple interactions
and the exploration of various conversation paths, given the same initial prompt. The _growing state_
continues until the tree reaches its _end state_, which is defined by a maximum number of messages or
other predetermined criteria.


Parameters within the data collection platform govern the behaviour of the tree state machine, such as
the average number of messages collected for each parent message or the maximum tree depth. These


4


parameters enable researchers to fine-tune the data collection process according to their specific
research goals and requirements, ensuring a more targeted and efficient approach to gathering data.
Parameters varied during the collection of the dataset. Current settings can be found in Appendix G.


In summary, the tree state machine is a structured and systematic method for managing the progression
of message trees during the data collection process. By guiding each tree through a series of states,
from initial prompt review to growing and reaching its _end state_, the tree state machine ensures
the collection of high-quality, diverse, and relevant data. Additionally, the inclusion of platform
parameters allows for the customization of the data collection process to align with specific research
objectives, further enhancing the effectiveness and utility of this approach.


**3.3** **Contributor Guidelines**


To achieve a high degree of quality and consistency across a wide range of contributors, we issue
clear and detailed guidelines. A full copy of these guidelines at the present time can be found in
Appendix A. Our guidelines follow three main goals: First, clarify the meanings, scales, and criteria
for assigning labels and rankings during the labelling and ranking tasks. Second, make assistant
responses be polite, helpful, concise, friendly, and safety-aware and third, instruct prompts and
prompter replies to explore a diverse and challenging set of inputs to the assistant role.


The guidelines establish a framework for safely interacting with an automated assistant by drawing
inspiration from the concept of _informed consent_ . Rather than categorically denying large parts of
request categories, we aim to provide the prompter with useful feedback, for example drawing special
awareness to dangerous activities, elaborating on weaknesses of automated assistants, such as hallucinations, and discouraging and denying requests asking for illegal or highly inappropriate content. In
our validation experiments in training assistant models based on OpenAssistant Conversations, we
observe a high degree of consistency of the trained models’ outputs with our given guidelines.


Although guideline adherence is already high in our models after training, our approach is completely
compatible with deploying additional safety measures during inference, such as secondary models to
filter or modify ill-suited user input.


**3.4** **Quality Control & Content Moderation**


We take a multi-pronged approach to quality assurance, with the main pillars being a system of
reward points & leaderboards, and manual review of flagged content by human moderators. This both
maximizes the quality of contributions, while effectively utilizing the limited time of the volunteer
moderators. In an effort to demonstrate progress and achievement to users, and to encourage highquality contributions, our system allocates points for the completion of tasks. These points contribute
to various leaderboards, including daily, weekly, monthly, and all-time rankings. A level system
also exists, wherein higher point accumulation results in elevated levels, reflecting veteran status and
engagement. In the future, this system could potentially be developed further to facilitate preferential
access to more engaging tasks or similar perks.


The distribution of points is contingent upon task type, as certain tasks require greater effort, such as
the _reply as assistant_ task (compared to the _create a prompt_ task). A significant portion of points
is deferred and reliant on interactions with other users. For instance, a user’s assistant reply may
gather many additional points if it is subsequently deemed non-spam and highly ranked by other
users. Inversely, points may be reduced or lost for answers that are labeled as spam or down-voted by
consensus of other users.


Within the moderator section of the website, an alternative leaderboard, designated the _Trollboard_,
is exhibited. This leaderboard assesses users based on an aggregate of negative labels, reports, and
down-votes received for their contributions. This approach enables human moderators to proactively
scrutinize potentially misbehaving users in a comprehensive manner. The Trollboard has proven to
be an effective tool in addressing the numerical disparity between users and moderators, maximizing
the collective efforts of contributors to identify undesirable contributions.


Users further have the option to report messages to moderators for manual review, either via the
platform, or directly via communication on a community chat server. Moderators have the ability to
delete individual messages, or all messages of a given user, at their own discretion. Deleted messages
are retained, but marked as deleted and are not exported for training.


5


**4** **Dataset Composition**


The full dataset consists of 161,443 messages (91,829 prompter and 69,614 assistant messages)
distributed across 66,497 conversation trees, in 35 different languages, annotated with 461,292
quality ratings. This includes 8,576 synthetic messages, leaving 152,867 human-submitted messages.
Of the 66,497 total conversation trees, we consider 10,968 complete, meaning the full number of
messages has been collected and the moderation process for these trees has been concluded. 52,159
incomplete trees are in the prompt lottery state, meaning they only consist of a single initial prompt.
The completed trees contain 92,365 messages.


The set of categories for which _Likert_ -scale human labels are collected is Creativity, Quality, Humor,
Helpfulness, Violence, and Rudeness. The set of categories for which binary human labels are
collected is Language Mismatch, Not Appropriate, Personally Identifiable Information, Hate Speech,
and Sexual Content. We additionally release the rank of each assistant message compared to other
assistant messages submitted for the same prompt, computed from the preference rankings of several
human annotators. To merge the rankings of multiple (possibly conflicting) annotators, we use a
variant of Tideman’s method, described in Appendix B.



































Figure 2: **Left:** Relative share of the most frequent languages in the dataset. **Right:** Distribution of
contributions among users.


The dataset is dominated by English and Spanish messages as illustrated in Figure 2 (left). The
prominence of English is expected as a result of the community around OpenAssistant originating
in the English-speaking open-source machine learning community. The high quantity of Spanish
messages can be attributed to the publicity given to OpenAssistant by prominent figures in the
Spanish machine learning community. Figure 2 (right) illustrates how a small number of power users
contributed a significant proportion of the dataset. This must be taken into account when considering
possible biases in the data. Although significant effort went into preventing responses directly copied
from other sources, it is possible that some users utilised automated techniques to enter data.


We release the dataset on the Hugging Face Hub [4] in several variants: One variant containing all
collected initial prompts, one variant containing all trees that are considered completed, one variant
containing all trees (completed and in-progress), and one variant containing messages filtered out as
spam. For most purposes, such as instruction-tuning of language models, we recommend using the
variant containing all completed trees, and include non-completed trees if more data is required.


**5** **Contributor Demographics and Satisfaction**


To gain a deeper understanding of the contributors’ demographics, a Google Form survey was sent out
as an announcement on the project’s Discord channel. The survey consists of 3 parts with questions
on demographic information, personal motivation and user satisfaction. At the time of the release of
this paper, a total of 270 survey answers have been collected. Results can be seen in Figures 7,8,9,10,
and Tables 3 and 4 (all in the Appendix). Respondent answers indicate a homogeneity towards
identifying as male, but strong diversity in the reported levels of education, stated motivations for
contribution, proficiency in artificial intelligence, and use cases for the technology. Over 95% of
respondents either agreed or strongly agreed with the statement "Overall, I’m glad I have contributed


4 `[https://huggingface.co/OpenAssistant/oasst1](https://huggingface.co/OpenAssistant/oasst1)`


6


to OpenAssistant." About 40% reported this being their first time contributing to an open-source
project. We note that the method of recruiting via Discord is biased towards users who are present on
the platform and have been active around the time of the announcement. Active contributors are also
expected to be more likely to respond.


**6** **Experimental Validation**


**6.1** **Instruction Tuning & Preference Modeling**


We focus on the development and evaluation of fine-tuned language models based on Pythia [3],
LLaMA [2], and Falcon [23]. Pythia and Falcon are state-of-the-art language models with permissive
open-source licenses, while LLaMA is a powerful language model with a bespoke non-commercial
license. Specifically we train supervised fine-tuned (SFT) models, reward models (RM [13],[24]),
and, using the trained reward models, reinforcement-learned models (RLHF) [5] .


Model **LMEH** **VEL** **OAIE** **HE**


gpt-3.5-turbo (ChatGPT) 1110 0.87 0.72
EleutherAI/pythia-12b 60.33
OpenAssistant/pythia-12b-sft-v8-7k-steps 60.28 997 0.10 0.10
tiiuae/falcon-40b 72.29
OpenAssistant/falcon-40b-sft-top1-560 74.04 1192 0.26 0.09
OpenAssistant/falcon-40b-sft-mix-1226 74.40 1053 0.44 0.13
huggyllama/llama-65b 67.24
OpenAssistant/oasst-sft-7e3-llama-30b 68.03 979 0.52 0.20
OpenAssistant/oasst-rlhf-3-llama-30b-5k-steps 68.51 1068 0.51 0.15

Table 1: Comparison of model evaluation scores on different LLM benchmarks: **LMEH:** lmevaluation-harness [25] (average scores, see online leaderboard for more details) **VEL:** Vicuna Elo
Rank [21] **OAIE:** OpenAI Evals [26] **HE:** HumanEval [27] (for all benchmarks, higher is better).
We have chosen to leave the Hugging Face Hub identifiers as the model names for identifiability.


Table 1 shows evaluation scores of a selection of baseline and trained models on a set of standard
benchmark datasets. Evaluations were performed externally using FastEval and evaluation results
are hosted on a leaderboard, which is continually being updated [6] . LMEH refers to the average
performance on a set of widely-used NLU tasks consisting of BoolQ, PIQA, HellaSwag, WinoGrande,
ARC-e, ARC-c and OBQA [7] . We omit instruction-centric experiments (VEL, OAIE, HE) for the
base models, as these benchmarks are unsuitable for non-instruction-tuned models. The results show
that models using OpenAssistant Conversations are consistently outperforming the corresponding
baseline models (in the case of LLaMA even a larger baseline model). RLHF outperforms SFT in
some benchmarks, but not in others. For Falcon-based models, _sft-top1_ is trained only on top-ranked
conversation threads, whereas _sft-mix_ mixes OpenAssistant Conversations with other instruction
datasets (details in Appendix H). The varied evaluation scores demonstrate that by combining different
data sources, the nature of the resulting model can be readily influenced. Ranks across benchmarks
are not consistent, which could indicate the unsuitability of automatic evaluations for language
models, or could indicate that different models and datasets lead to different capabilities. The results
also show that while open-source models are close to matching ChatGPT in some benchmarks, others
still show large performance gaps. Anecdotally, users report OpenAssistant models to be less robotic
and more human-sounding than commercial models and report generations to have high quality and
diversity in domains such as creative writing, conversational messaging, and drafting social media
posts.


5All trained models are released at `[https://huggingface.co/OpenAssistant](https://huggingface.co/OpenAssistant)`
6 `[https://github.com/FastEval/FastEval](https://github.com/FastEval/FastEval)`, `[https://tju01.github.io/ilm-eval/](https://tju01.github.io/ilm-eval/)`
7For readability, Table 1 contains aggregated LMEH scores. Details in the online leaderboard.


7


**6.2** **Spam and Toxicity**


To understand the concordance between human and automated toxicity detection, we employ toxicity
detection methods based on Detoxify [28] to obtain automated ratings for six distinct categories,
classifying whether a message is toxic, obscene, threatening, insulting, attacking a certain identity or
sexually explicit. We limit our analysis to those languages that are supported by the toxicity detection
method, covering English, Spanish, Russian, French, and Italian. These languages represent the
majority of OASST1 messages (over 83%).



Using automated toxicity ratings, we are able
to systematically assess the correlation between
these ratings and human-assigned toxicity labels
(hate speech, not appropriate, and sexual content). Based on a sample of 115,153 messages,
we compute the correlation between automatic
and human-annotated toxicity labels, which is
visualized in Figure 3. We see a correlation between human and automatic labels in at least
one element of each row and column of the correlation matrix, suggesting agreement between
human annotators and off-the-shelf toxicity detection models. The results serve to both validate the capabilities and show limitations of
AI-driven toxicity detection in comparison to
human judgement and may inform future work
in this area.



Figure 3: Correlation between human labels and
Detoxify outputs for all messages in Detoxifysupported languages.



In addition to analysing the correlation between human-assigned toxicity labels and automated
ratings, we extend the application of the Detoxify model to assess the efficacy of the moderation
process for the same languages described earlier. To facilitate this analysis, we define two categories
of messages: _deleted_ messages, which encompass those that either failed to pass the community
moderation process or were subsequently manually removed by moderators, and _retained_ messages,
which successfully made it through to the dataset. In order to provide a comprehensive evaluation
of the moderation process, we calculated average values for each of the six Detoxify categories for
both _deleted_ and _retained_ messages. The values obtained for this analysis are based on a sample of
74,781 messages. We excluded messages in trees that were incomplete at the time of export, as these
messages may be subject to removal by the moderation process.


Our analysis, presented in Table 2 shows that the values for all six toxicity categories are markedly
higher for _deleted_ messages compared to _retained_ messages. This significant difference demonstrates
the effectiveness of the moderation processes in place, as messages removed from the dataset are on
average rated as significantly more toxic by the Detoxify model than messages allowed to remain in
the dataset.


Toxicity Obscene Threat Insult Identity Attack Explicit N
State


Deleted 4.625% 1.965% 0.411% 2.085% 0.651% 1.39% 3422
Retained 0.988% 0.574% 0.102% 0.715% 0.121% 0.177% 71359

Table 2: Detoxify outputs across six categories of toxicity, comparing _deleted_ and _retained_ messages.


While _deleted_ messages are rated as more toxic than _retained_ messages by the Detoxify model across
all categories, the average toxicity values for these messages are still small. This implies toxicity
ratings from models like Detoxify alone are not sufficient to determine when messages are unsuitable
for inclusion in the dataset. Reasons for deleting non-toxic messages may include a lack of factual
accuracy, or poor grammar. Additionally, messages which are children of deleted messages must
themselves be deleted even if they appear to be acceptable in isolation.


8


**7** **Limitations**


**Reward model data collection.** InstructGPT [13] trained reward models on ranking data of messages
generated by their initial SFT model, while our reward models are trained using ranking data of
human-generated messages. We chose to do so because we were already collecting this ranking data
as part of our general efforts, for use in quality control, spam filtering, and dataset sub-sampling.
While subjectively, many users report our RLHF models to follow instructions more closely (also
compare Section 6.1), the models do not deliver the same uniform and significant improvements over
SFT models as reported in [13]. We hypothesize that the difference in data collection for the reward
model could at least partially explain this gap. We plan to collect ranking data based on our own SFT
models in the future to verify these assumptions. Further research is necessary to determine more
precise criteria for collecting data useful to RLHF.


**Subjective, Cultural, and Contribution Frequency Biases.** The open nature of our project introduces a unique set of challenges when it comes to controlling for biases within the dataset. Annotators
from diverse backgrounds contribute to the dataset, with demographics that are simultaneously heterogeneous in some dimensions and homogeneous in others (see Section 5). Specifically, 89.1% of
the annotators identify as male, with a median age of 26. This demographic profile may inadvertently
introduce biases in the dataset, as it is bound to reflect the values, perspectives, and interests of
the annotators. (We pose that some of this could be mitigated by introducing more constrained
conversations, for example sampling a random Wikipedia page to determine the conversation topic.)
Further, users’ participation levels differ significantly. More engaged users contribute a greater
number of annotations (see Figure 2), which likely leads to over-representation of their values and
interests in the dataset. Consequently, the dataset may not adequately capture the diverse perspectives
that a more balanced distribution of contributions could have provided. Further research is necessary
to determine the effect of uneven contributor distributions have when given a clear, general task.


**Possibility of Unsafe Content.** While we have implemented measures to detect and remove harmful
messages, our system is not infallible. It is possible that the dataset still contains unsafe content. We
believe that the open nature of the project allows for data filtering to be conducted in a transparent
manner, ultimately converging on the highest possible standards. Nevertheless, the potential presence
of residual unsafe content in the dataset necessitates careful evaluation of any models trained on it.


Given the limitations discussed above, we advocate for the use of our models in academic research
contexts only. We strongly encourage researchers to thoroughly investigate the safety and bias of any
model before employing it in downstream tasks. The released models may exhibit unsafe behavior
and are likely susceptible to prompt injection attacks. The alignment of LLMs is a crucial aspect of
AI research, and we hope that our contributions can help advance the field of AI alignment. However,
we also acknowledge that current alignment techniques are not perfect and can even exacerbate
certain biases [29]. As such, researchers should exercise caution when using these models and be
cognizant of their limitations. Additionally, it is essential to continue refining alignment techniques
and advancing the field of AI alignment in order to mitigate these limitations and develop more
reliable and robust LLMs.


**8** **Safety and Ethical Implications**


Large language models are prone to generating inaccurate information about people, places, or facts,
commonly known as ‘hallucinations’ [30, 31]. LLMs can also produce toxic or hateful content and
fail to follow provided constraints [32]. Additionally, these models tend to incorporate biases present
in their training data, leading to unfair and discriminatory outputs [33]. While methods such as
_RLHF_ can mitigate some of these shortcomings, they may exacerbate others [34, 29]. We hope that
alignment methods using OpenAssistant Conversations can fix some of these issues [13], but we
acknowledge that achieving full alignment is a complex and ongoing challenge.


We recognize that sufficiently powerful language models can have a significant impact on society [35],
and therefore we believe it is essential to promote transparency in their development and deployment.
OpenAssistant Conversations is our contribution to this goal of transparency.


9


**Acknowledgments and Disclosure of Funding**


Our greatest thanks go to the many volunteer contributors, of human data, code, moderation, documentation, and community organization. Absent of any financial incentives, this project is a stunning
and unprecedented display of global cooperation of humans for the purpose of advancing and democratizing AI research. In addition, several organizations have contributed to this project with resources:
Redmond AI provided training compute. Stability AI and Hugging Face provided inference compute.
We thank Olivier Dehaene at Hugging Face for close collaboration and personal support. Weights
& Biases provided their full MLOps solution to the entire team. LAION provided legal input and
acts as the website addressee. We thank Til Jasper Ullrich for running LLM benchmark evaluations,
Luke Thomas Kaiser for running evaluations on model bias, and Silvia Pareti for detailed feedback
on the manuscript. Agradecemos a Carlos Santana por la promoción a gran escala de la plataforma
de recopilación de datos para la comunidad hispanohablante. [8]


**References**


[1] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. _Advances in neural information_
_processing systems_, 30, 2017.


[2] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open
and efficient foundation language models. _arXiv preprint arXiv:2302.13971_, 2023.


[3] Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O’Brien, Eric
Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff,
et al. Pythia: A suite for analyzing large language models across training and scaling. _arXiv_
_preprint arXiv:2304.01373_, 2023.


[4] Iason Gabriel. Artificial intelligence, values, and alignment. _Minds and Machines_, 30(3):411–
437, September 2020.


[5] Yufei Wang, Wanjun Zhong, Liangyou Li, Fei Mi, Xingshan Zeng, Wenyong Huang, Lifeng
Shang, Xin Jiang, and Qun Liu. Aligning large language models with human: A survey, 2023.


[6] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma,
Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath,
Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny
Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine
Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann,
and Jared Kaplan. Training a helpful and harmless assistant with reinforcement learning from
human feedback, 2022.


[7] Kawin Ethayarajh, Yejin Choi, and Swabha Swayamdipta. Understanding dataset difficulty with
_V_ -usable information. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari,
Gang Niu, and Sivan Sabato, editors, _Proceedings of the 39th International Conference on_
_Machine Learning_, volume 162 of _Proceedings of Machine Learning Research_, pages 5988–
6008. PMLR, 17–23 Jul 2022.


[8] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha,
Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee,
Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun,
Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts,
Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch,
Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel
Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen,
Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina,
Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm,


8Translated from English by OpenAssistant, first try


10


Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise AgueraArcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. Lamda: Language models for dialog
applications, 2022.


[9] Jacob Hilton, R Nakano, S Balaji, and John Schulman. Webgpt: Improving the factual accuracy
of language models through web browsing. _OpenAI Blog, December_, 16, 2021.


[10] Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving, et al. Teaching
language models to support answers with verified quotes. _arXiv preprint arXiv:2203.11147_,
2022.


[11] Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei,
Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences,
2020.


[12] Ratanond Koonchanok, Ya-Chen Pan, and Hyeju Jang. Tracking public attitudes toward chatgpt
on twitter using sentiment analysis and topic modeling. _ArXiv_, abs/2306.12951, 2023.


[13] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin,
Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton,
Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano,
Jan Leike, and Ryan Lowe. Training language models to follow instructions with human
feedback, 2022.


[14] Leandro von Werra, Jonathan Tow, reciprocated, Shahbuland Matiana, Alex Havrilla, cat
state, Louis Castricato, Alan, Duy V. Phung, Ayush Thakur, Alexey Bukhtiyarov, aaronrmm,
Fabrizio Milo, Daniel, Daniel King, Dong Shin, Ethan Kim, Justin Wei, Manuel Romero, Nicky
Pochinkov, Omar Sanseviero, Reshinth Adithyan, Sherman Siu, Thomas Simonini, Vladimir
Blagojevic, Xu Song, Zack Witten, alexandremuzio, and crumb. CarperAI/trlx: v0.6.0: LLaMa
(Alpaca), Benchmark Util, T5 ILQL, Tests, March 2023.


[15] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal
policy optimization algorithms. _arXiv preprint arXiv:1707.06347_, 2017.


[16] Lukas Budach, Moritz Feuerpfeil, Nina Ihde, Andrea Nathansen, Nele Noack, Hendrik Patzlaff, Felix Naumann, and Hazar Harmouch. The effects of data quality on machine learning
performance, 2022.


[17] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi,
and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. _arXiv preprint arXiv:2212.10560_, 2022.


[18] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy
Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model.
`[https://github.com/tatsu-lab/stanford_alpaca](https://github.com/tatsu-lab/stanford_alpaca)`, 2023.


[19] Yuvanesh Anand, Zach Nussbaum, Brandon Duderstadt, Benjamin Schmidt, and Andriy Mulyar.
Gpt4all: Training an assistant-style chatbot with large scale data distillation from gpt-3.5-turbo.
`[https://github.com/nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all)`, 2023.


[20] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning
with gpt-4. _arXiv preprint arXiv:2304.03277_, 2023.


[21] Vicuna. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality. `[https:](https://vicuna.lmsys.org/)`
`[//vicuna.lmsys.org/](https://vicuna.lmsys.org/)`, 2023.


[22] Herbert H. Clark and Susan E. Brennan. Grounding in communication. In Lauren Resnick,
Levine B., M. John, Stephanie Teasley, and D., editors, _Perspectives on Socially Shared_
_Cognition_, pages 13–1991. American Psychological Association, 1991.


[23] Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra
Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic,
Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. Falcon-40B: an open large
language model with state-of-the-art performance, 2023.


11


[24] Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy
Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, Nelson Elhage, Zac Hatfield-Dodds,
Danny Hernandez, Jackson Kernion, Kamal Ndousse, Catherine Olsson, Dario Amodei, Tom
Brown, Jack Clark, Sam McCandlish, Chris Olah, and Jared Kaplan. A general language
assistant as a laboratory for alignment, 2021.


[25] Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence
Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric
Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language
model evaluation, September 2021.


[26] OpenAI. Gpt-4 technical report, 2023.


[27] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto,
Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul
Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke
Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad
Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias
Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex
Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain,
William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra,
Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer,
Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech
Zaremba. Evaluating large language models trained on code, 2021.


[28] Laura Hanu and Unitary team. Detoxify. Github. https://github.com/unitaryai/detoxify, 2020.


[29] Amelia Glaese, Nat McAleese, Maja Tr˛ebacz, John Aslanides, Vlad Firoiu, Timo Ewalds, Maribeth Rauh, Laura Weidinger, Martin Chadwick, Phoebe Thacker, Lucy Campbell-Gillingham,
Jonathan Uesato, Po-Sen Huang, Ramona Comanescu, Fan Yang, Abigail See, Sumanth
Dathathri, Rory Greig, Charlie Chen, Doug Fritz, Jaume Sanchez Elias, Richard Green, Soˇna
Mokrá, Nicholas Fernando, Boxi Wu, Rachel Foley, Susannah Young, Iason Gabriel, William
Isaac, John Mellor, Demis Hassabis, Koray Kavukcuoglu, Lisa Anne Hendricks, and Geoffrey
Irving. Improving alignment of dialogue agents via targeted human judgements, 2022.


[30] Wojciech Kry´sci´nski, Bryan McCann, Caiming Xiong, and Richard Socher. Evaluating the
factual consistency of abstractive text summarization. _arXiv preprint arXiv:1910.12840_, 2019.


[31] Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. Retrieval augmentation reduces hallucination in conversation. _arXiv preprint arXiv:2104.07567_, 2021.


[32] Peter Henderson, Koustuv Sinha, Nicolas Angelard-Gontier, Nan Rosemary Ke, Genevieve
Fried, Ryan Lowe, and Joelle Pineau. Ethical challenges in data-driven dialogue systems. In
_Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society_, pages 123–129,
2018.


[33] Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. Measuring
and mitigating unintended bias in text classification. In _Proceedings of the 2018 AAAI/ACM_
_Conference on AI, Ethics, and Society_, pages 67–73, 2018.


[34] Ethan Perez, Sam Ringer, Kamile Lukoši˙ ut¯ e, Karina Nguyen, Edwin Chen, Scott Heiner, Craig˙
Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kadavath, Andy Jones, Anna Chen, Ben
Mann, Brian Israel, Bryan Seethor, Cameron McKinnon, Christopher Olah, Da Yan, Daniela
Amodei, Dario Amodei, Dawn Drain, Dustin Li, Eli Tran-Johnson, Guro Khundadze, Jackson
Kernion, James Landis, Jamie Kerr, Jared Mueller, Jeeyoon Hyun, Joshua Landau, Kamal
Ndousse, Landon Goldberg, Liane Lovitt, Martin Lucas, Michael Sellitto, Miranda Zhang,
Neerav Kingsland, Nelson Elhage, Nicholas Joseph, Noemí Mercado, Nova DasSarma, Oliver
Rausch, Robin Larson, Sam McCandlish, Scott Johnston, Shauna Kravec, Sheer El Showk,
Tamera Lanham, Timothy Telleen-Lawton, Tom Brown, Tom Henighan, Tristan Hume, Yuntao
Bai, Zac Hatfield-Dodds, Jack Clark, Samuel R. Bowman, Amanda Askell, Roger Grosse, Danny
Hernandez, Deep Ganguli, Evan Hubinger, Nicholas Schiefer, and Jared Kaplan. Discovering
language model behaviors with model-written evaluations, 2022.


12


[35] Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang,
Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, et al. Ethical and social risks of
harm from language models. _arXiv preprint arXiv:2112.04359_, 2021.


[36] Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei.
Deep reinforcement learning from human preferences, 2023.


[37] Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec
Radford, Dario Amodei, and Paul Christiano. Learning to summarize from human feedback,
2022.


[38] T. N. Tideman. Independence of clones as a criterion for voting rules. _Social Choice and_
_Welfare_, 4(3):185–206, September 1987.


[39] David Rozado. The political biases of chatgpt. _Social Sciences_, 12(3):148, Mar 2023.


13


# **Appendix**

#### **Table of Contents**

**A Contributor Guidelines** **14**


**B** **Ranking merging** **18**


**C Web application for data collection** **19**


**D Filtering ChatGPT Inputs** **22**


**E** **Online Survey Results** **22**


**F** **Word Cloud** **27**


**G Collection parameters** **27**


**H Training Configuration** **29**


**I** **Evaluation Tasks** **30**


**J** **Political Compass Evaluations** **30**


**K Community Engagement** **31**


**L** **Dataset Documentation (Data card)** **34**


**M Author Statement** **34**


**A** **Contributor Guidelines**


We provide the guidelines presented to the users for the creation of the dataset.

```
# Guidelines

Below is a list of guidelines that should be adhered to for each possible task
available when building the dataset. To see some examples of how the guidelines
can be applied, visit the examples document.

Please consider checking out our survey
[here](https://forms.gle/vBW7b2kMzjCoehkH9). You can use it to rate each
guideline and leave feedback for each task.

If you have further suggestions to improve any of our guidelines, or want to add
more examples, create a pull request or suggest them on our
[GitHub](https://github.com/LAION-AI/Open-Assistant).

## 1. General rules

- Always make sure to read and understand the guidelines to each task before
 fulfilling it.
- Try to follow the guidelines as closely as possible.
- If you are unsure whether a message violates a guidelines, contact us at our
 Discord.

```

14


```
- Use the thumbs-up/thumbs-down system to further mark messages that are of high
 or low quality.

## 2. Providing an assistant reply {#assistant-reply}

### Do:

- Remain polite and treat the user with respect, even when not given the same
 courtesy.
- Talk in a friendly and approachable manner, unless specifically requested
 otherwise.
- Present only information that has been verified by credible sources that can
 be backed up, unless specifically requested otherwise.
- Make sure the user is aware when given unverified information.
- Inform the user about the potential dangers when being asked for advice
 regarding a topic with high risk, such as medicine, law or chemistry.
- When being asked about a high-risk topic, make sure the user knows that as a
 language model, the assistant is susceptible to producing incorrect
 information, and that no actions should be taken regarding the assistant reply
 without the opinion of a professional.
- When being asked to give an opinion as the default persona of the assistant,
 make sure to bring up at least 2 common viewpoints and ensure that these
 aren’t expressed as the opinions of the assistant.
 - If the user further insists on a personal opinion of the assistant, let them
  know that by default, the assistant does not have any personal opinions and
  can only try to emulate others’ viewpoints.
- Ask for clarification if it’s unclear what the user is asking for.
- Use paragraphs and line breaks to make larger replies more readable.
- Make use of [Markdown syntax](https://www.markdownguide.org/basic-syntax) to
 better format lists, tables or blocks of code.
 - If you are using a codeblock to write code in a particular language, specify
  it to enable
  [syntax highlighting]
  (https://www.markdownguide.org/extended-syntax/#syntax-highlighting).
  You can find all supported abbreviations
  [here](https://github.com/jincheng9/markdown_supported_languages
  #heres-a-full-list-of-supported-languages).
- Be consistent in the style and tone of the assistant.

### Don’t:

- Copy and paste text from other sources without editing. **This includes
 ChatGPT.**
- Supply text that violates the law of Germany, UK, USA, or your country of
 residence.
- Write content encouraging:
 - Violence
 - Violation of the rights of a third party
 - Pedophilia
- Provide the user with information that could be used for self-harm if there is
 plausible suspicion of intent to self-harm.
- Provide personal information of third parties that isn’t publicly available.
- Ask for personal information unless it is relevant to the issue and can’t be
 used to determine the identity of the user, such as country of residence or
 occupation. The user should be allowed to refuse to give up any information.
- Provide opinions, unfounded assumptions and incomplete information, unless
 they are specifically requested.
- Purposefully curate information to guide the conclusion, i.e. don’t hide facts
 to present a particular narrative.

```

15


```
- Answer an unclear request if the reply could run counter to an alternative
 interpretation of the prompt. Ask the user to elaborate or rephrase instead.
- Dodge a question, unless it violates a guideline.
- Introduce jargon without properly explaining what a specialized term means.
 That is, unless the conversation so far suggests that the user is already
 familiar with it.
- Leave typos or grammatical errors in the assistant replies, unless
 specifically requested to do so.
- Overload the user with too much information. Keep replies concise, but include
 further details that relate to and expand upon the user’s request.
- Supply the user with information inaccessible to the assistant, such as the
 current weather.
- Reply in a language different from the one intended for the dataset, unless
 specifically requested to do so.

## 3. Providing an initial prompt or user reply {#user-reply}

### Do:

- Ask questions that reflect real-life situations and needs.
- Ask questions that might be directed towards search engines or specialists.
- Make requests that encourage lateral thinking and/or require specialized
 knowledge.
- Use a mix between questions that are straightforward and questions without a
 clear answer.
- Introduce a variety in prompts by using different phrasing, degrees of
 politeness or amount of context given.
- Consider the previous replies and prompts that lead up to the current one.
- Try to build upon the topic and ask a sensible follow-up question when
 replying to the assistant.

### Don’t:

- Write prompts without a clear request.
- Supply text that violates the law of Germany, UK, USA, or your country of
 residence.
- Make requests that override the original purpose of the assistant, i.e.
 jailbreak the model.
- Make requests that leave the assistant with no other choice but to refuse in
 order to avoid the generation of harmful content.
- Submit a prompt similar or identical to a prompt you previously submitted.
- Change the topic of a conversation without prefacing it accordingly when
 replying to the assistant.
- Leave typos and grammatical errors in the prompt.
- Reply in a language different from the one intended for the dataset, unless
 the context of the conversation requires it.

## 4. Classifying an assistant reply {#classifying-assistant}

### Do:

- Rate every criteria of each reply, unless it can’t be discerned because it is
 spam or inappropriate.
- Judge quality based on how well the reply adheres to the guidelines. Factual
 accuracy and helpfulness are first and foremost.
- Make sure to read the reply thoroughly.
- Use the [label explanations](#label-explanation) to determine which labels
 apply to the reply.
- Research to make sure whether the reply is factually accurate.

```

16


```
- Skip a classification if you are unable to determine the validity of reply.

### Don’t:

- Judge quality based on personal beliefs. Assuming an opinion was warranted,
 fulfills the users request and doesn’t violate any guidelines, it should not
 impact the rating of the reply.
- Skip a label just because the reply is spam. Each label can help the model
 improve.
- Rate a reply if you are unsure if it factually accurate or satisfies the
 request of the user.

## 5. Classifying an initial prompt or user reply {#classifying-user}

### Do:

- Rate every criteria of each prompt, unless it can’t be discerned because it is
 spam or inappropriate.
- Judge quality based on how well the prompt adheres to the guidelines.
- Make sure to read the prompt thoroughly.
- Use the [label explanations](#label-explanation) to determine which labels
 apply to the prompt.

### Don’t:

- Judge quality based on personal beliefs. The opinion of the user should not
 impact the rating of the prompt.
- Skip a label just because the reply is spam. Each label can help the model
 improve.

## 6. Ranking assistant replies {#ranking-assistant}

### Do:

- Make sure to read every available reply.
- Think about which reply best satisfies the request of the user.
- Rank replies based on how well they adhere to the guidelines. Factual accuracy
 and helpfulness are first and foremost.
- Penalize replies that fail to provide adequate warnings or caveats.
- Penalize replies that are difficult to read due to a lack of formatting,
 capitalization or other errors.
- Penalize replies if the requested information is obfuscated by superfluous
 details that make up a large part of the message.
- Rank replies that admit to not knowing the answer below factually correct, but
 above factually incorrect replies.

### Don’t:

- Rank replies based on personal beliefs. Assuming an opinion was warranted,
 fulfills the users request and doesn’t violate any guidelines, it should not
 impact the rating of the reply.
- Rank replies based on how long and short they are - instead, find out which
 reply best answers the query of the user.

## 7. Label Explanation {#label-explanation}

**Is the message spam?**

Is the message an unrelated piece of text that doesn’t contribute to the

```

17


```
previous prompt in any way? If it’s an initial prompt, is it something unlikely
to be used to initiate a conversation with the assistant?

**Is it a bad reply, as an answer to the prompt task?**

Does the assistant reply fail to engage the request of the user or does it
contain incorrect information? Does it ignore relevant parts of the prompt?

### Tags

**Not Target Language**: Unless specifically requested, is the language used
different from the target language of the dataset?

**Not Appropriate**: In case of a user prompt, is it missing a clear request?

**Contains PII**: Does the message contain personally identifying information
(PII) that isn’t publicly available and can be used to determine the identity of
the user or a private third party?

**Hate Speech**: Does the message contain prejudice towards a group classified
as protected according to the law of Germany, UK, USA or your country of
residence?

**Sexual Content**: Does the message contain or request sexually explicit
content, such as descriptions of sexual acts?

### Scales

**Low Quality/High Quality**: How well does the message adhere to the
guidelines? For assistant replies, factual accuracy and helpfulness are first
and foremost.

**Serious/Humorous**: Does the message contain sarcasm, a play on words or other
humorous embellishments?

**Ordinary/Creative**: How unlikely is the question to be asked? Is it
challenging the assistant with a problem that requires creative thinking? Does
the question include lots of context? On the assistant’s side, is the reply
straightforward, or does it contain a unique approach to a problem?

**Rude/Polite**: How polite is the message towards the recipient? Does it use
words like "please" or is it dismissive, rude or insulting towards the other
side of the conversation?

**Violent/Harmless**: Does the message contain descriptions of explicit violent
acts? Is it glorifying, encouraging or downplaying violent behavior?

```

**B** **Ranking merging**


Reinforcement learning from human feedback [36, 37] comprises a set of techniques that all aim
to optimize the output distribution of a language model using the preference structure provided by
human rankers. To get a preference structure that is well aligned to users, we cannot just rely on the
opinions of individual rankers, due to the high variance in human preferences. Since our objective
is to collect data for a generally capable digital assistant, every ranking of possible responses is
performed by K independent rankers, (see, section 3.1). Once this is done, we need to fuse these K
individual opinions into one consensus opinion usable in training preference models.


18


We perform this preference fusion by treating it as a ranked-voting problem, whose objective it
is to maintain the preferences as faithfully as possible. The method chosen for this is known as
"ranked pairs" or "Tideman’s method" [38]. Simplified, this method creates a sorted list of "winners"
according to the strength of the preference of one element over the others. The way the preference
strength is measured is by considering all preference pairs in the input votes: For example, if the
votes are two times _A > B > C_ and one time _B > A > C_, this would mean that the pair _A > B_
exists 2 times, while _A > C_ exists three times. The method then orders the winners by winning
strength, i.e. here _A > C_ would happen before _A > B_, and constructs a directed graph using the
preferences, i.e. _A > C_ would become an edge _A →_ _C_ . Edges are added one by one according to
their weight (higher first), as long as no cycle is produced. If the edge would produce a cycle, it is
skipped as the corresponding dominant preferences are already in the graph.


The resulting directed acyclic graph can be turned into a preference structure by recursively removing
source nodes and adding them to the back of a preference list, since source nodes are, by construction,
not preferred over any other item left in the graph. In practice, one can speed up the construction by
not explicitly constructing the preference graph and fusing the graph con- and destruction.


**C** **Web application for data collection**


Figures 4, 5, and 6 show the data collection platform UI interface for labelling, reply input, and
ranking, respectively. The UI also allows the user to choose a random task type, the task is automatically chosen according the previously discussed heuristics, which allows for more variation and
engagement.


The UI did grow and change over time as the project developed, so the experience of early users was
less refined as shown here.


19


Figure 4: A preview of the page for labelling tasks. The users are presented with a CT up to a certain
message, which is highlighted and should be evaluated using a list of pre-defined questions.


20


Figure 5: A preview of the page for ranking assistant replies. Users are provided with CT up to a
message, and a couple of responses that should be ranked according to how good they answer the
given message. The interaction is drag & drop based. Additionally, users can choose to mark all
answers as factually incorrect.


Figure 6: A preview of the page for replying as an assistant. Users are provided with a CT up to a
prompter message, and they should provide a response to it. In this example, the CT contains only
one message. Users can use Markdown in their responses. An additional progress bar is added below
the input field to incentivize longer messages.


21


**D** **Filtering ChatGPT Inputs**


The widespread use of ChatGPT at the time of our data collection meant that some of the inputs
could be performed using ChatGPT, rather than being fully created by humans. Our guidelines
specifically target this, explicitly discouraging contributors from copy-pasting responses generated
by other AI models. Users who were found to post ChatGPT-generated responses were banned,
and their contributions were deleted. Furthermore, we used multiple automatic tests to catch such
cases. For instance, we searched for and removed messages that contained text such as "as a large
language model" or "knowledge cutoff after September 2021". Moreover, users were encouraged
to up-vote and down-vote responses they came across from other users, which also helped weeding
out low-quality, generic, (possibly AI generated) responses. We note that these mechanisms cannot
remove all ChatGPT-generated content, and more post-hoc filtering may be necessary.


**E** **Online Survey Results**


We asked users to provide feedback based on the overall experience, while taking part in the data
collection process. Results are provided in Tables 3, 4 and Fig. 7, 8, 9, 10.


Figure 7: Demography of 270 respondents


22


It was clear from the website what to do and how to contribute messages to the training data.
Strongly agree 33.70%
Agree 47.04%
Neither agree nor disagree 11.48%
Disagree 6.67%
Strongly disagree 1.11%
I’ve felt I could always ask for help from the community and the moderators.
Strongly agree 30.74%
Agree 27.78%
Neither agree nor disagree 31.85%
Disagree 7.41%
Strongly disagree 2.22%
I found the tasks enjoyable and engaging.
Strongly agree 20.00%
Agree 42.22%
Neither agree nor disagree 27.78%
Disagree 8.15%
Strongly disagree 1.85%
I found the tasks repetitive.
Strongly agree 11.48%
Agree 30.00%
Neither agree nor disagree 37.04%
Disagree 18.15%
Strongly disagree 3.33%
While doing rating or ranking tasks, I found the messages to be really high quality.
Strongly agree 16.67%
Agree 42.59%
Neither agree nor disagree 30.00%
Disagree 9.26%
Strongly disagree 1.48%
Overall, I’m glad I have contributed to OpenAssistant.
Strongly agree 81.11%
Agree 14.44%
Neither agree nor disagree 2.96%
Disagree 0.74%
Strongly disagree 0.74%
Table 3: User Satisfaction Survey


Have you contributed to other community projects besides this one?


No, this is my first time contributing 111
Yes I have contributed to a few projects 110
Yes, I have contributed to multiple open source projects 44
Prefer not to say 5

Table 4: Previous Contributions


23


Figure 8: Familiarity with AI based on level of education. The average values are shown as outlines
in the bar chart.


24


Figure 9: Motivations for contribution based on level of education. The average values are shown as
outlines in the bar chart.


25


Figure 10: Planned personal use-cases for OpenAssistant. The average values are shown as outlines
in the bar chart.


26


**F** **Word Cloud**


We present some statistics on the words used the most in Fig. 11.


Figure 11: Word clouds for 33 topics extracted from the English subset of the open-assistant dataset.
The number of topics was selected by selecting the peak coherence from 40 LDA models. One can
observe high topic variety from biology (second top row), cooking (top row), and music (second to
last row), all the way up to quantum physics (bottom row).


**G** **Collection parameters**


We provide more information on the parameters used to collect new data samples.


27


Parameter value


max active trees 100
max initial prompt review 100
max tree depth 5
max children count 2
num prompter replies 1
goal tree size 9
num reviews initial prompt 3
num reviews reply 3
auto mod enabled true
auto mod max skip reply 25
auto mod red flags 4
p full labeling review prompt 1
p full labeling review reply assistant 1
p full labeling review reply prompter 0.1
acceptance threshold initial prompt 0.6
acceptance threshold reply 0.6
num required rankings 3
p activate backlog tree 0.1
min active rankings per lang 20
lonely children count 2
p lonely child extension 0.75
recent tasks span sec 300
max pending tasks per user 8
max prompt lottery waiting 1000

Table 5: Collection parameters


_max active trees_ : Maximum number of concurrently active message trees in the database. No new
initial prompt tasks are handed out to users if this number is reached


_max initial prompt review_ : Maximum number of initial prompts under review before no more initial
prompt tasks will be handed out.


_max tree depth_ : Maximum depth of message tree.


_max children count_ : Maximum number of reply messages per tree node.


_num prompter replies_ : Number of prompter replies to collect per assistant reply.


_goal tree size_ : Total number of messages to gather per tree.


_num reviews initial prompt_ : Number of peer-review checks to collect in the ‘INITIAL_PROMPT_REVIEW‘ state


_num reviews reply_ : Number of peer review checks to collect per reply (other than initial prompt).


_auto mod enabled_ : Flag to enable/disable auto moderation.


_auto mod max skip reply_ : Automatically set tree state to ‘halted_by_moderator‘ when more than the
specified number of users skip replying to a message. (auto moderation)


_auto mod red flags_ : Delete messages that receive more than this number of red flags if it is a reply or
set the tree to ‘aborted_low_grade‘ when a prompt is flagged. (auto moderation)


_p full labeling review prompt_ : Probability of full text-labeling (instead of mandatory only) for initial
prompts.


_p full labeling review reply assistant_ : Probability of full text-labeling (instead of mandatory only) for
assistant replies.


_p full labeling review reply prompter_ : Probability of full text-labeling (instead of mandatory only) for
prompter replies.


_acceptance threshold initial prompt_ : Threshold for accepting an initial prompt.


_acceptance threshold reply_ : Threshold for accepting a reply.


28


_num required rankings_ : Number of rankings in which the message participated.

_p activate backlog tree_ : Probability to activate a message tree in BACKLOG_RANKING state when
another tree enters a terminal state.

_min active rankings per lang_ : When the number of active ranking tasks is below this value when a
tree enters a terminal state an available trees in BACKLOG_RANKING will be activated (i.e. enters
the RANKING state).

_lonely children count_ : Number of children below which parents are preferred during sampling for
reply tasks.

_recent tasks span sec_ : Time in seconds of recent tasks to consider for exclusion during task selection.

_max pending tasks per user_ : Maximum number of pending tasks (neither canceled nor completed) by
a single user within the time span defined by ‘recent_tasks_span_sec‘.

_max prompt lottery waiting_ : Maximum number of prompts in prompt_lottery_waiting state per
language. If this value is exceeded no new initial prompt tasks for that language are generated.


**H** **Training Configuration**


Following [13] and as introduced in Section 1, we train supervised fine-tuned models ( _SFT_ ), reward
models (RM), and a PPO fine-tuned models based on RM’s predictions. We use as base models the
popular decoder-only Pythia [3] and LLaMA [2].


**Conversation format** We sample threads in the CTs and provide them as input text to the model by
using some additional special tokens. More specifically, a thread composed of prompts (P) P1 _,_ P2 _, . . ._
and replies (R) R1 _,_ R2 _, . . ._ is provided as input to the model with the following format:


<prompter_token> P1 <endoftext_token> <assistant_token> R1 <endoftext_token>
<prompter_token> P2 <endoftext_token> <assistant_token> R2 <endoftext_token>
_. . ._


Each of the prompts and the replies consists potentially of multiple tokens after tokenizing.


_**SFT-mix**_ **details.** The _sft-mix_ training data configuration used for training _OpenAssistant/falcon-_
_40b-sft-mix-1226_ is a mixture of OASST1 with other instruction tuning datasets, according to the
following configuration:


sft9−stage2:
# oasst_export: 100.00% (29899)
# vicuna: 50.00% (16963)
# code_alpaca: 50.00% (9510)
# oa_wiki_qa_bart_10000row: 100.00% (9434)
# grade_school_math_instructions: 100.00% (8351)
# dolly15k: 100.00% (14250)


use_custom_sampler: true
datasets:
−oasst_export:
lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk" # sft−8.0
input_file_path: 2023−06−02_oasst_all_labels.jsonl.gz
val_split: 0.05
top_k: 2
−vicuna:
fraction: 0.5
val_split: 0.025
max_val_set: 250
−code_alpaca:
fraction: 0.5
val_split: 0.05
max_val_set: 250
−oa_wiki_qa_bart_10000row:


29


val_split: 0.05
max_val_set: 250
−grade_school_math_instructions:
val_split: 0.05
−dolly15k:
val_split: 0.05
max_val_set: 300


More details can be found on the Hugging Face hub and in our open-source codebase.


**Supervised fine-tuning.** During this phase, we fine-tune pretrained models for the regular language
modelling tasks based on our conversational data. We mask tokens that correspond to prompts and
only train to predict tokens that correspond to assistant replies.


**Reward model.** For the reward model training, we replace the language modelling head with a
linear layer producing a single output _rθ_, corresponding to the predicted score for the last reply of
the conversation. We use replies to the same prompt and their rankings as described in Appendix B.
Following [13], assuming _K_ distinct replies, we produce - _K_ 2 - comparisons and train to minimize the
loss

loss( _θ_ ) = _−_ ~~�~~ _K_ [1] ~~�~~ _E_ ( _x,yw,yl_ )[log( _σ_ ( _rθ_ ( _x, yw_ ) _−_ _rθ_ ( _x, yl_ )))] _,_
2


where _σ_ is the sigmoid function and _yw_ corresponds to a preferred completion for the pair of _yw_ and
_yr_ . We also optionally add another regularization parameter that prevents the predicted values from
diverging too much. Performance is measured by measuring the ability to predict the better reply
among pairs of replies with different rank, on a held-out validation set.


**PPO training.** We fine-tune the _SFT_ model by producing assistant replies to unanswered questions.
We use the RM to score these replies and train with PPO, using the trlx framework [9] . Following [13],
we also add a per-token KL penalty from the _SFT_ model at each token to avoid instability and
over-optimization to the RM model.


All details and current training parameters are publicly available under `[https://github.com/](https://github.com/LAION-AI/Open-Assistant/tree/main/model/model_training)`
`[LAION-AI/Open-Assistant/tree/main/model/model_training](https://github.com/LAION-AI/Open-Assistant/tree/main/model/model_training)` .


**I** **Evaluation Tasks**


This section aims to provide a brief explanation of the evaluation tasks used in Table 1.


_lm-evaluation-harness [25]_ is a framework for evaluating language models on a variety of standardized tasks. We focus on the subset of BoolQ, PIQA, HellaSwag, WinoGrande, ARC-e, ARC-c and
OBQA.


_Vicuna Elo Rank [21]_ leverages LLMs to judge other LLMs and assign them a ranking according to
the ELO system.


_OpenAI Evals [26]_ is a set of general LLM evaluation benchmarks OpenAI released along with the
announcement of GPT-4, and to which OpenAI has asked the community to contribute.


_HumanEval [27]_ is a benchmark OpenAI released along with the announcement of Codex and sets
out to measure "measure functional correctness for synthesizing programs from docstrings".


**J** **Political Compass Evaluations**


The political leanings of ChatGPT have been investigated in [39]. We evaluated a model fine-tuned on
OpenAssistant Conversations on a subset of the given tests. Prompts were standardized and multiple
samples were drawn, with majority vote deciding on the final answer for each question. Figure 12
depicts the result. We stress that these are very preliminary results and should not be taken with


9 `[https://github.com/CarperAI/trlx](https://github.com/CarperAI/trlx)`


30


large degrees of certainty, as the community has yet to find consensus on the exact methodology to
perform such evaluations. We will update this section with improved results in the future. Our limited,
preliminary results show the model trained on OpenAssistant Conversations to be more balanced and
varied in its political leanings than ChatGPT.


Figure 12: Comparison of evaluations on test for political leanings. For original ChatGPT results and
references to tests used, see [39]


**K** **Community Engagement**


Throughout the collection of OpenAssistant Conversations, a large global community has been built,
including an active Discord group, and a GitHub repository with over 200 contributors.


Figure 13 shows the growth of the Discord community throughout the duration of data collection.


Figure 14 shows new commits to the GitHub repository over time.


Figure 15 shows the growth in stars on the GitHub repository over time.


Figure 16 shows popularity of OpenAssistant by YouTube’s videos’ views on the theme over time.


These Figures serve as a strong reminder of what can be achieved by the collective effort of many
volunteers, even in a field where research has thus far been largely monopolized by a small number
of industrial labs.


In addition, by comparing the massive influx of new contributors and subscribers to the emergence of
Open Assistant themed videos, it shows how certain media events have influenced the development.


31


Figure 13: Discord users in the OpenAssistant group over time.


Figure 14: GitHub commits to the OpenAssistant repository over time.


32


Figure 15: GitHub stars on the OpenAssistant repository over time.


Figure 16: YouTube’s videos’ views on the OpenAssistant theme over time.


33


**L** **Dataset Documentation (Data card)**


Dataset documentation can be found at `[https://huggingface.co/OpenAssistant/oasst1](https://huggingface.co/OpenAssistant/oasst1)` .


**M** **Author Statement**


We confirm that we bear all responsibility in case of any violation of rights during the collection of
the data or other work, and will take appropriate action when needed, e.g. to remove data with such
issues.


34