Can we trust the evaluation on ChatGPT?

May combine two claims.

On the P-Stance dataset, we observe a 0.74-point improvement in the micro-F scores and a 0.26 point in the macro-F scores when comparing Feb 13 ChatGPT plus to V1.

“On the P-Stance dataset, we observe a 0.74-point improvement in the micro-F scores and a 0.26 point in the macro-F scores when comparing Feb 13 ChatGPT plus to V1.”

• AssumptionThe authors state this as an assumption: Zhang et al., 2022 used either the November 30 or December 15 version of ChatGPT (V1) to obtain their results.
• AssumptionThe authors state this as an assumption: In Zhang et al., 2022, F1m is the micro average and F1-avg is the macro average.

May combine two claims.

There is a 6.1 point improvement on the micro-F and 1.89 point improvement on the macro-F when comparing Feb 13 ChatGPT plus to V1 on the SemEval task.

“We see a smaller but nonnegligible improvement-6.1 point on the micro-F and 1.89 point on the macro-F-when comparing Feb 13 ChatGPT plus to V1 on the same task.”

• AssumptionThe authors state this as an assumption: Zhang et al., 2022 used either the November 30 or December 15 version of ChatGPT (V1) to obtain their results.
• AssumptionThe authors state this as an assumption: In Zhang et al., 2022, F1m is the micro average and F1-avg is the macro average.

May combine two claims.

Even though the performance has dropped, it is still quite an improvement compared to V1.

“Even though the performance has dropped, it is still quite an improvement compared to V1.”

• AssumptionThe authors state this as an assumption: Zhang et al., 2022 used either the November 30 or December 15 version of ChatGPT (V1) to obtain their results.
• AssumptionThe authors state this as an assumption: In Zhang et al., 2022, F1m is the micro average and F1-avg is the macro average.

Classified as probably not a substantial claim.

We cannot exclude the possibility that ground-truth output answers/labels of NLP tasks were submitted to the platform and ChatGPT was trained with such data.

“Although we do not have any documented evidence that the ground-truth output answers/labels of the NLP tasks were submitted to the platform and the ChatGPT model has been trained with such data, we cannot exclude such possibility either.”

May combine two claims.

ChatGPT is constantly updated, partly based on human inputs from more than 100 million users via RLHF.

“Given that the ChatGPT's training datasets are unknown and that ChatGPT is constantly updated, partly based on human inputs from more than 100 million users via Reinforcement Learning from Human Feedback (RLHF) (OpenAI, 2022), it is impossible to ascertain the lack of data leakage, especially for the datasets that have been on the internet.”

• cites OpenAI, 2022Not checked yet

  Not checked in the free run.

May combine two claims.

ChatGPT is a closed model without information about its training dataset and how it is currently being trained.

“However, given that the ChatGPT is a closed model without information about its training dataset and how it is currently being trained, there is a large loxodonta mammal in the room: how can we know whether ChatGPT has not been contaminated with the evaluation datasets?”

• cites OpenAI, 2022Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

It is impossible to rule out the possibility of contamination without the model owners' in-depth analysis.

“Although it is not possible to definitely answer these questions, it is also impossible to rule out the possibility of contamination without the model owners' in-depth analysis.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

OpenAI has been updating the model primarily to address the model's problematic behaviors by making it more restricted, which led to the observation, although largely anecdotal, that the model has become 'less impressive.'

“However, we would also like to note that OpenAI has been updating the model primarily to address the model's problematic behaviors by making it more restricted, which led to the observation, although largely anecdotal, that the model has become 'less impressive.'”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Due to the same reasons mentioned above, we could not try multiple queries for each input and could not estimate the uncertainty of the performance.

“Due to the same reasons mentioned above, we could not try multiple queries for each input and could not estimate the uncertainty of the performance.”

Classified as probably not a substantial claim. May combine two claims.

We used an open-source API for experiments with Feb 13 ChatGPT plus, which opened a new chat session per query.

“In contrast, we used an open-source API for our experiments with the Feb 13 ChatGPT plus version, which opened a new chat session per query.”

Probably not a substantial claim: may not pass S2 (checkable).

As long as the trend of closed models and continuous training loop continues, it will become more challenging to prevent data leakage and ensure fair evaluation of models.

“As long as the trend of closed models and continuous training loop continues, it will become more challenging to prevent data leakage (training-test data contamination) and ensure fair evaluation of models.”

Classified as probably not a substantial claim. May combine two claims.

Although we cannot rule out the explanation that ChatGPT is simply superior to previous models, it is also impossible to rule out the possibility of data leakage.

“Although we cannot rule out the explanation that the ChatGPT is simply superior to previous models, it is also impossible to rule out the possibility of data leakage.”

Probably not a substantial claim: may not pass S2 (checkable).

Given results showing that even benign contamination can lead to measurable differences, making claims about these models' zero-shot or few-shot inference capabilities require a more careful inspection of the training datasets of these models.

“Given the results showing that even benign contamination can lead to measurable differences, making claims about these models' zero-shot or few-shot inference capabilities require a more careful inspection of the training datasets of these models.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

This may be one explanation for the drop in performance between Jan 30 and Feb 13 observed in Table 1 but recent work showed this to have an insignificant effect, although on a different dataset.

“This may be one explanation for the drop in performance between Jan 30 and Feb 13 observed in Table 1 but recent work showed this to have an insignificant effect, although on a different dataset (Kocoń et al., 2023).”

• cites Kocoń et al., 2023Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Newer versions of ChatGPT are more likely to be exposed to past queries to the platform.

“newer versions of ChatGPT are more likely to be exposed to past queries to the platform”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Due to rate limitations, the collection was not done in one sitting since Jan 30 ChatGPT did not entertain more than approx 40 queries in an hour.

“However, due to the rate limitations, this was not done in one sitting since Jan 30 ChatGPT did not entertain more than a fixed (approx.40) queries in an hour.”

Classified as probably not a substantial claim.

In each sitting, a single chat session was used to feed multiple inputs, one at a time, to ChatGPT, which may have accumulated context for subsequent inputs.

“Additionally, in each sitting, a single chat session was used to feed multiple inputs, one at a time, to ChatGPT 10 , which may have accumulated context for subsequent inputs.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Given that ChatGPT is based on InstructGPT3 in which some NLP dataset contamination was already documented, this raises further concerns if V1 too may have been contaminated.

“Given that ChatGPT is based on InstructGPT3 in which some NLP dataset contamination was already documented (Brown et al., 2020), this raises further concerns if V1 too may have been contaminated.”

• cites Brown et al., 2020Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S1 (asserted).

Recent studies have claimed that ChatGPT outperforms most previous models for stance detection on existing evaluation datasets such as SemEval 2016 Task6 and P-stance, even in a zero-shot setting.

“Recent studies have claimed that ChatGPT outperforms most of the previous models proposed for this task (Zhang et al., 2022) on a few existing evaluation datasets, such as the SemEval 2016 Task6 dataset (Mohammad et al., 2016(Mohammad et al., , 2017) ) and P-stance (Li et al., 2021), even in a zero-shot setting where the model was not fine-tuned on the task-specific training data.”

• cites Zhang et al., 2022Not checked yet

  Not checked in the free run.

• cites Mohammad et al., 2016Not checked yet

  Not checked in the free run.

• cites Li et al., 2021Not checked yet

  Not checked in the free run.