arxiv_figs_cleaned

The arxiv_figs__x family of the ElliotVL supervised-fine-tuning pool, after VLM cleaning.

images	86,252
QA turns	690,365
answers rewritten by the cleaning pass	0
QA created by the cleaning pass (new_qa)	not measured for this family
shards	10

How this was cleaned

A vision-language model read each image together with its QA and judged the item. The pass is not a filter that only removes rows — it rewrites answers it finds wrong but salvageable, drops what it cannot salvage, and adds QA where the image supports more than the source provided. Each row carries the judge's own record in clean_meta, including the cleaner identity, the policy it applied, and its per-item scores for legibility, richness and coverage.

A large share of the QA here was written by the cleaning pass, not by the original dataset. Across the pool that share runs from roughly half to over 80% of a family's turns, and it is reported in the table above. Those rows have an empty org_answer because no pre-clean original exists. Treat them as model-generated supervision: they were judged against the image, but they are not human annotation, and model-written QA is where formatting defects are most likely.

The effect on the answers that were carried over is substantive rather than cosmetic. In this pool the pass corrects values, not just wording — a curve's critical point restated from 4.00 to 2.00, a computed ratio from 1 to Approximately 1.33 — and for some families it removed the content entirely, which is why those families are absent here rather than published empty.

answer vs org_answer

• answer — the cleaned answer. Train on this.
• org_answer — the pre-cleaning answer from the same (image, question) in the uncleaned pool. It is empty for QA the cleaning pass added, which have no original.

For this family org_answer is empty throughout: recovering it means scanning the whole pre-clean family, which is raw family is 7 GB (> 6 GB cap). The cleaned answer is unaffected.

Where the two differ, the difference is the correction. Keeping both makes every edit auditable instead of asking you to take the cleaning on trust.

This dataset supersedes an earlier upload of the same family that was built from the pre-clean pool. That version carried the uncorrected answers and has been deleted.

Schema

• image — HF Image(), renders directly in the dataset viewer
• image_sha256 — content hash of the image
• n_qa — number of turns attached to this image
• qa — list of {question, answer, org_answer}, in source order
• family / clean_meta — provenance and the judge's per-row record

A row with more than one entry in qa is a conversation over a single image, not a set of independent questions — the turns depend on each other and the image belongs to all of them. Keep them together and in order when training.

Parquet row groups are written at ~32 MB so the dataset viewer pages without stalling.

What is NOT claimed

No rejection sampling was run against this family: there is no accept/reject verdict per row, and answer is the cleaned reference rather than a model generation that a string verifier confirmed. For that, use the *_RS_think / *_rejected datasets in this org.

Examples

10 rows taken straight from this dataset. Images are the original files as stored here - no downscaling, no recompression - and every turn is shown in full.

Example 1 — 5 turns on one image

#	question	answer
1	What is the average execution time of the Long-term scheduler in the 50 trials?	The average execution time of the Long-term scheduler in the 50 trials is 1815.166 seconds.
2	Which scheduler has the shortest average execution time, and what is that time?	The Short-term scheduler has the shortest average execution time, which is 0.056 seconds.
3	How does the average execution time of the Mid-term scheduler compare to that of the Long-term scheduler?	The average execution time of the Mid-term scheduler (58.347 seconds) is significantly shorter than that of the Long-term scheduler (1815.166 seconds).
4	What are the headers for the two columns in the table?	The headers are 'Scheduler' and 'Execution Time (s)'.
5	List the three schedulers shown in the table.	Long-term scheduler, Mid-term scheduler, and Short-term scheduler.

Example 2 — 7 turns on one image

#	question	answer
1	Which model achieved the highest score in the Montezuma’s Revenge environment, and what was the score?	The SND-V model achieved the highest score in the Montezuma’s Revenge environment with a score of 21,565.
2	In which environments did the SND-VIC model achieve the highest scores, and what were those scores?	The SND-VIC model achieved the highest scores in the following environments: - Gravitar with a score of 6,712. - Venture with a score of 2,188. - Private Eye with a score of 17,313.
3	Which environment proved to be the most challenging for all models, and what was the outcome?	The Pitfall environment proved to be the most challenging for all models, as none of the models managed to achieve any score.
4	Compare the performance of the Baseline model and the SND-STD model in the Solaris environment. Which model performed better and by how much?	In the Solaris environment, the SND-STD model performed better than the Baseline model. The SND-STD model achieved a score of 12,460, while the Baseline model achieved a score of 12,344. The SND-STD model outperformed the Baseline model by 116 points.
5	What was the score of the RND model in the Private Eye environment?	14996
6	Which model achieved the lowest score in the Montezuma's Revenge environment, and what was that score?	The Baseline model achieved the lowest score with 400.
7	What was the score of the SND-VIC model in the Montezuma's Revenge environment?	7838

Example 3 — 8 turns on one image

#	question	answer
1	What is the range of final values for the fraction of the total cluster mass in second-generation stars ((M_{\rm 2G}/M_{\rm tot})) according to the table?	The final values of the fraction of the total cluster mass in second-generation stars ((M_{\rm 2G}/M_{\rm tot})) range from 0.53 to 0.80.
2	Which model has the highest final number of stars ((N_f)) and what is its corresponding (M_{\rm 2G}/M_{\rm tot}) value?	The model "4m-sg025c20wf" has the highest final number of stars ($N_f$) with 1,727,911 stars, and its corresponding $M_{\rm 2G}/M_{\rm tot}$ value is 0.54.
3	Identify the model with the highest fraction of second-generation stars ((M_{\rm 2G}/M_{\rm tot})) and provide its final number of stars ((N_f)).	The model "sg025c20w04sf" has the highest fraction of second-generation stars ((M_{\rm 2G}/M_{\rm tot})) with a value of 0.80, and its final number of stars ((N_f)) is 145,363.
4	How does the final number of stars ((N_f)) compare between the models "sg01c20sf" and "sg01c20wf"?	The final number of stars ((N_f)) for the model "sg01c20sf" is 47,381, whereas for the model "sg01c20wf" it is 111,517. The model "sg01c20wf" has a significantly higher final number of stars compared to "sg01c20sf".
5	What is the final fraction of second-generation stars ((M_{\rm 2G}/M_{\rm tot})) for the model "4m-sg025c20sf" and how does it compare to the model "4m-sg025c20wf"?	The final fraction of second-generation stars ((M_{\rm 2G}/M_{\rm tot})) for the model "4m-sg025c20sf" is 0.57, while for the model "4m-sg025c20wf" it is 0.54. The model "4m-sg025c20sf" has a slightly higher fraction of second-generation stars compared to "4m-sg025c20wf".
6	Which model actually has the highest final number of stars ($N_f$) and what is that value?	The model "4m-sg025c20wf" has the highest final number of stars ($N_f$) with 1,727,911.
7	What is the final number of stars ($N_f$) and the fraction of second-generation stars ($M_{\rm 2G}/M_{\rm tot}$) for the model "sg025c20w04wf"?	The model "sg025c20w04wf" has a final number of stars ($N_f$) of 239,498 and a fraction of second-generation stars ($M_{\rm 2G}/M_{\rm tot}$) of 0.77.
8	What is the lowest final number of stars ($N_f$) listed in the table and which model corresponds to it?	The lowest final number of stars is 47,381, corresponding to the model "sg01c20sf".

Example 4 — 8 turns on one image

#	question	answer
1	Which model variant achieves the highest BLEU score for the 1st sentence at B@4, and what is the score?	The "full" model achieves the highest BLEU score for the 1st sentence at B@4, with a score of 5.52.
2	How does the METEOR score for the 2nd sentence compare between the "no context" and "full" models?	The METEOR score for the 2nd sentence is higher in the "full" model (8.28) compared to the "no context" model (7.79).
3	What is the CIDEr score for the 3rd sentence in the "online" model, and how does it compare to the "no context" model?	The CIDEr score for the 3rd sentence in the "online" model is 19.36, which is the same as the score in the "no context" model (19.36).
4	Which model variant shows the most significant improvement in BLEU scores from the 1st to the 2nd sentence?	The 'full' model shows the most significant decrease in BLEU scores from the 1st to the 2nd sentence, with B@1 decreasing from 26.33 to 21.46, B@2 from 13.98 to 9.06, B@3 from 8.45 to 4.40, and B@4 from 5.52 to 2.33.
5	What is the trend observed in the CIDEr scores across the 1st, 2nd, and 3rd sentences for the "no context" model?	In the "no context" model, the CIDEr scores show a decreasing trend across the 1st, 2nd, and 3rd sentences: 31.56 for the 1st sentence, 19.37 for the 2nd sentence, and 19.36 for the 3rd sentence.
6	What are the BLEU scores (B@1, B@2, B@3, B@4) for the 'online' model on the 1st sentence?	The scores are 24.93, 12.38, 7.45, and 4.77 respectively.
7	What is the CIDEr score for the 'full' model on the 2nd sentence?	The CIDEr score is 20.17.
8	What are the column headers listed in the table?	The column headers are B@1, B@2, B@3, B@4, M, and C.

Example 5 — 9 turns on one image

#	question	answer
1	What is the neutral pressure (p_n) for a collisionality parameter (\tau_c) of 10?	The neutral pressure $p_n$ for a collisionality parameter $\tau_c$ of 10 is 1.
2	How does the normalized ion flow velocity (|v_f|/c_s) change as the collisionality parameter (\tau_c) increases from 1 to 100?	As the collisionality parameter (\tau_c) increases from 1 to 100, the normalized ion flow velocity (|v_f|/c_s) increases from 0.3 to 1.9.
3	For which collisionality parameter (\tau_c) is the normalized height (z/\lambda_{De}) the smallest, and what is its value?	The normalized height (z/\lambda_{De}) is the smallest for the collisionality parameter (\tau_c) of 100, and its value is 8.
4	Compare the normalized ion density (n_i/n_{40}) for (\tau_c = 1) and (\tau_c = 100). How does it change?	The normalized ion density (n_i/n_{40}) increases from 0.1 for (\tau_c = 1) to 0.4 for (\tau_c = 100).
5	What is the relationship between the collisionality parameter (\tau_c) and the neutral pressure (p_n) based on the given data?	Based on the given data, as the collisionality parameter $\tau_c$ increases, the neutral pressure $p_n$ decreases. Specifically, $p_n$ is 10 for $\tau_c = 1$, 1 for $\tau_c = 10$, and 0.1 for $\tau_c = 100$.
6	What is the value of the normalized height $z/\lambda_{De}$ for a collisionality parameter $\tau_c$ of 10?	10
7	What is the value of the normalized ion flow velocity $|v_f|/c_s$ for a collisionality parameter $\tau_c$ of 10?	1.3
8	What is the value of the normalized ion density $n_i/n_{40}$ for a collisionality parameter $\tau_c$ of 10?	0.2
9	What is the value of the normalized height $z/\lambda_{De}$ for a collisionality parameter $\tau_c$ of 1?	12

Example 6 — 7 turns on one image

#	question	answer
1	Which model achieved the highest F1 score on the Hausa language dataset?	The Afro-xlmr-large model achieved the highest F1 score of 0.82.
2	What is the development loss of the Afro-XLM-R-base model, and how does it compare to the other models?	The development loss of the Afro-XLM-R-base model is 0.53, which is the lowest among all the models evaluated in the table.
3	How does the performance of the XLM-R-large model compare to the Afro-XLM-R-large model in terms of F1 score and development loss?	The XLM-R-large model recorded an F1 score of 0.7 and a development loss of 0.69, whereas the Afro-XLM-R-large model achieved a higher F1 score of 0.82 and a lower development loss of 0.56, indicating superior performance.
4	Which model has the highest development loss, and what is its corresponding F1 score?	The Afro-XLM-R-small model has the highest development loss of 0.75, with a corresponding F1 score of 0.77.
5	What do the results in the table suggest about the effectiveness of pre-trained models in African languages, particularly for sentiment analysis in low-resource languages like Hausa?	The table suggests that the Afro-XLM-R variants generally outperform the standard XLM-R-large model, with the Afro-xlmr-large model achieving the highest F1 score and the Afro-xlmr-base model achieving the lowest development loss.
6	What are the F1 score and dev loss for the Afro-xlmr-mini model?	The Afro-xlmr-mini model has an F1 score of 0.71 and a dev loss of 0.69.
7	Which numerical values in the table are formatted in bold text?	The F1 score of 0.82 and the dev loss of 0.53 are bolded.

Example 7 — 7 turns on one image

#	question	answer
1	Which region within Stephan's Quintet exhibits the highest shock velocity, and what is the corresponding mass flow and cooling time for that region?	Ridge 3 exhibits the highest shock velocity of 40 km s^-1. The corresponding mass flow for this region is 1.74 x 10^2 M_sun yr^-1, and the cooling time is 7.76 x 10^3 years.
2	What is the warm H({}{2}) mass for the Bridge region, and how does it compare to the warm H({}{2}) mass in Ridge 3?	The warm H2 mass for the Bridge region is 9.13 x 10^8 M_sun. Ridge 3 has two warm H2 mass values: 1.17 x 10^8 M_sun (at 13 km/s) and 1.35 x 10^6 M_sun (at 40 km/s). The Bridge mass is significantly higher than both values for Ridge 3.
3	What are the shock velocities and cooling times for the SQ-A region, and how do they relate to the derived warm H({}_{2}) masses?	The shock velocity for the SQ-A region is 5 km s^-1. The corresponding cooling time is 8.09 x 10^4 years. The derived warm H2 mass is 1.19 x 10^9 M_sun.
4	Which region has the shortest cooling time, and what is the warm H({}_{2}) mass for that region?	Ridge 3 has the shortest cooling time of 7.76 x 10^3 years. The warm H2 mass for this region is 1.35 x 10^6 M_sun.
5	What are the shock velocities, mass flows, cooling times, and warm H2 masses for the two entries of Ridge 2?	For Ridge 2, the first entry has a shock velocity of 6 km s^-1, mass flow of 6.61 x 10^3 M_sun yr^-1, cooling time of 6.30 x 10^4 yr, and warm H2 mass of 4.16 x 10^8 M_sun. The second entry has a shock velocity of 35 km s^-1, mass flow of 5.40 x 10^2 M_sun yr^-1, cooling time of 9.77 x 10^3 yr, and warm H2 mass of 5.27 x 10^6 M_sun.
6	What are the mass flow, cooling time, and warm H2 mass for Ridge 1 at a shock velocity of 8 km s^-1?	At a shock velocity of 8 km s^-1, Ridge 1 has a mass flow of 6.28 x 10^3 M_sun yr^-1, a cooling time of 5.33 x 10^4 yr, and a warm H2 mass of 3.35 x 10^8 M_sun.
7	What are the mass flow, cooling time, and warm H2 mass for the SQ-A region?	For the SQ-A region, the mass flow is 1.47 x 10^4 M_sun yr^-1, the cooling time is 8.09 x 10^4 yr, and the warm H2 mass is 1.19 x 10^9 M_sun.

Example 8 — 7 turns on one image

#	question	answer
1	What is the objective function for the Koza-2 symbolic regression problem, and how many variables does it involve?	The objective function for the Koza-2 symbolic regression problem is (x^5 - 2x^3 + x), and it involves 1 variable.
2	Describe the training set used for the Pagie-1 symbolic regression problem.	The training set used for the Pagie-1 symbolic regression problem is denoted as E[-5,5,0.4], which refers to a grid of points evenly spaced with an interval of 0.4, from -5 to 5 inclusive.
3	How many uniform random samples are drawn from the interval [-1, 1] for the training sets of Koza-2 and Koza-3 problems?	For both the Koza-2 and Koza-3 problems, 20 uniform random samples are drawn from the interval [-1, 1] for their training sets.
4	What is the difference in the number of variables between the Koza-3 and Pagie-1 symbolic regression problems?	The Koza-3 symbolic regression problem involves 1 variable, while the Pagie-1 symbolic regression problem involves 2 variables. Therefore, the difference in the number of variables between the Koza-3 and Pagie-1 problems is 1 variable.
5	What is the objective function for the Koza-3 symbolic regression problem?	The objective function for the Koza-3 problem is x^6 - 2x^4 + x^2.
6	What is the objective function for the Pagie-1 symbolic regression problem?	The objective function for the Pagie-1 problem is 1/(1+x^-4) + 1/(1+y^-4).
7	What is the specific notation used for the training set of the Koza-2 and Koza-3 problems?	The notation used is U[-1,1,20].

Example 9 — 6 turns on one image

#	question	answer
1	What is the number of genera and types for the quaternion algebra with discriminant (d(\mathfrak{A}) = 2)?	For the quaternion algebra with discriminant (d(\mathfrak{A}) = 2), there are 6 genera and 10 types.
2	How many genera and types are there for the quaternion algebra with discriminant (d(\mathfrak{A}) = 3)?	For the quaternion algebra with discriminant (d(\mathfrak{A}) = 3), there are 3 genera and 3 types.
3	Which discriminant value has the highest number of types, and what is that number?	The discriminant value (d(\mathfrak{A}) = 2) has the highest number of types, which is 10.
4	Compare the number of genera and types for discriminant values (d(\mathfrak{A}) = 1) and (d(\mathfrak{A}) = 6).	For (d(\mathfrak{A}) = 1), there are 9 genera and 9 types. For (d(\mathfrak{A}) = 6), there are 2 genera and 2 types. Thus, (d(\mathfrak{A}) = 1) has more genera and types compared to (d(\mathfrak{A}) = 6).
5	What are the headers for the three columns in the table?	The column headers are $d(\mathfrak{A})$, $#genera$, and $#types$.
6	For which values of $d(\mathfrak{A})$ is the number of genera equal to the number of types?	The number of genera is equal to the number of types for $d(\mathfrak{A}) = 1$, $d(\mathfrak{A}) = 3$, and $d(\mathfrak{A}) = 6$.

Example 10 — 8 turns on one image

#	question	answer
1	Which model variant achieves the highest Exact Match (EM) score on the Alipay dataset?	The complete RMRC model achieves the highest Exact Match (EM) score with an EM score of 9.28.
2	What is the F1 score of the RMRC model without the reward function ( r_b )?	The F1 score of the RMRC model without the reward function ( r_b ) is 54.13.
3	How does the performance of the RMRC model change when answer filtering is removed?	When answer filtering is removed (RMRC w/o Answer Filtering), the model's performance drops significantly to an EM score of 2.10 and an F1 score of 35.05.
4	Which model variant achieves the highest F1 score, and what is that score?	The complete RMRC model achieves the highest F1 score, which is 55.67.
5	Compare the performance of the baseline model BERT-S with the complete RMRC model in terms of EM and F1 scores.	The baseline model BERT-S achieves an EM score of 1.81 and an F1 score of 32.62. In comparison, the complete RMRC model achieves significantly higher scores with an EM score of 9.28 and an F1 score of 55.67.
6	What are the EM and F1 scores for the model variant 'RMRC w/o Question Fusing'?	The EM score is 8.03 and the F1 score is 53.71.
7	What are the EM and F1 scores for the model variant 'RMRC w/o QS Training'?	The EM score is 8.52 and the F1 score is 51.72.
8	What are the EM and F1 scores for the model variant 'RMRC w/ Confidence-based Selector'?	The EM score is 8.76 and the F1 score is 52.34.