---
license: cc-by-nc-4.0
pretty_name: MatSciChartQ-Traces (CC BY-NC 4.0)
task_categories:
  - visual-question-answering
  - question-answering
language:
  - en
tags:
  - chart-understanding
  - materials-science
  - electrochemistry
  - battery-aging
  - battery-formation
  - polymer-electrolyte
  - reasoning-traces
  - chain-of-thought
  - vision-language-model
  - numerical-extraction
  - non-commercial
size_categories:
  - 1K<n<10K
configs:
  - config_name: aging_matrix_heatmap
    data_files: questions/aging_matrix_heatmap.parquet
  - config_name: aging_mode_trajectory_zoo
    data_files: questions/aging_mode_trajectory_zoo.parquet
  - config_name: aging_trajectory_metric_histograms
    data_files: questions/aging_trajectory_metric_histograms.parquet
  - config_name: conductivity_molarity_curve
    data_files: questions/conductivity_molarity_curve.parquet
  - config_name: electrode_soc_quadrant_scatter
    data_files: questions/electrode_soc_quadrant_scatter.parquet
  - config_name: fade_trajectory_protocol_overlay
    data_files: questions/fade_trajectory_protocol_overlay.parquet
  - config_name: formation_protocol_fade_band
    data_files: questions/formation_protocol_fade_band.parquet
  - config_name: halfcell_simulated_full_cell_overlay
    data_files: questions/halfcell_simulated_full_cell_overlay.parquet
  - config_name: partial_capacity_vs_rpt_scatter
    data_files: questions/partial_capacity_vs_rpt_scatter.parquet
  - config_name: predicted_vs_measured_cycle_life_parity
    data_files: questions/predicted_vs_measured_cycle_life_parity.parquet
  - config_name: protocol_parameter_swarm_efc
    data_files: questions/protocol_parameter_swarm_efc.parquet
  - config_name: trajectory_metric_correlation_heatmap
    data_files: questions/trajectory_metric_correlation_heatmap.parquet
---
# MatSciChartQ-Traces (CC BY-NC 4.0)

MatSciChartQ-Traces pairs every question in the MatSciChartQ benchmark (`translorentz/materials-science-visual-bench-nc`, version 1.8.1-nc) with a step-by-step reasoning trace that works from the chart pixels to the graded answer. There are 4,037 traces covering all twelve chart families and all 40 question types of the parent benchmark, one trace per question row, each shipped alongside the chart image and the exact final answer. The charts themselves span battery ageing, battery formation and polymer-electrolyte behaviour, re-rendered as publication-style figures from published research datasets, and the questions ask for the quantities a working scientist would read off them, such as end-of-life crossings, medians across the plotted set of cells (the cohort), fit slopes and counts of strongly correlated metric pairs.

The defining property of this library is that the traces are programmatically exact. The parent benchmark's rendering pipeline is the single source of the numbers behind every chart, so each trace is produced by a deterministic template whose only inputs are the instance ground truth and intermediate quantities recomputed by the same numerics code that produced the graded answer. No language model wrote or filled any part of any trace, no step was produced by reading pixels back off an image, and no value was sampled from a model or invented. Every number that appears in a trace is either a true ground-truth value or a true intermediate from the pipeline's own computation, and where a step describes reading a value off the chart, the wording hedges it as an estimate while the number itself is the exact pipeline value. Presentation facts, such as which panels a variant draws, how the curves are coloured, and whether a residual axis is signed or logarithmic, are likewise derived from the instance's recorded render settings rather than assumed by the template. This is the deliberate contrast with distilled reasoning datasets, in which the reasoning is sampled from a model and then filtered. Reliability here is a property of construction rather than of post-hoc verification.

## Chart families

The family names below are code identifiers shared with the parent benchmark and keep their original spelling. Each row of the table says what the family's charts plot and which upstream dataset the numerics come from. The twelve Parquet files total about 57 MB, and every row is self-contained, with its chart image embedded alongside the text columns.

| config | traces | what the charts plot | source |
|---|---:|---|---|
| `aging_matrix_heatmap` | 128 | signed heatmap of how charging-protocol parameters impact mechanistic state-of-health metrics | van Vlijmen 2025 |
| `aging_mode_trajectory_zoo` | 512 | per-cell trajectories of degradation modes (lithium-inventory loss and active-material losses) over cycling | Che 2025 |
| `aging_trajectory_metric_histograms` | 368 | histogram panels of ageing-trajectory metrics across a cell cohort | van Vlijmen 2025 |
| `conductivity_molarity_curve` | 103 | polymer-electrolyte ionic conductivity and lithium diffusivity against salt molality | Xie 2023 |
| `electrode_soc_quadrant_scatter` | 512 | scatter of electrode state-of-charge shifts, with quadrants separating degradation modes | van Vlijmen 2025 |
| `fade_trajectory_protocol_overlay` | 640 | capacity-fade curves against equivalent full cycles, overlaid by cycling protocol | van Vlijmen 2025 |
| `formation_protocol_fade_band` | 768 | capacity-fade trajectories of formation-protocol cohorts against cycle number | Cui 2024 |
| `halfcell_simulated_full_cell_overlay` | 251 | measured full-cell voltage curves overlaid with half-cell-based simulations, plus residuals | Cui 2024 |
| `partial_capacity_vs_rpt_scatter` | 276 | partial-charge capacity features against reference-performance-test capacity | Che 2025 |
| `predicted_vs_measured_cycle_life_parity` | 256 | parity scatter of predicted against measured cell-life metrics | Che 2025 |
| `protocol_parameter_swarm_efc` | 159 | swarm plots of an ageing metric against charging-protocol stressor bins | van Vlijmen 2025 |
| `trajectory_metric_correlation_heatmap` | 64 | Pearson correlation matrix of ageing-trajectory metrics | van Vlijmen 2025 |

## What a trace looks like

Each trace is written as natural-language prose in four to seven steps that mirror how a careful reader would work. An orientation step identifies the chart type, the axes and the cohort. Two or more read steps extract concrete quantities from the plotted data, for example the position of a dashed end-of-life threshold, the endpoints of the steepest fade trajectory, or the per-bin midlines of a swarm plot, in which every cell appears as one point within its parameter bin. One or more reasoning steps then combine the reads through the materials-science operation the question requires, such as a cohort median, a least-squares slope converted to per-thousand-cycle units, a linear interpolation to a threshold crossing, a quartile boundary, or a count of strongly correlated metric pairs on a correlation matrix, at the magnitude threshold the question states. Where a read locates a value against an axis, the trace first states the axis's drawn tick pitch, so the calibration a pixel reader needs is explicit. Where the answer admits a genuinely independent second route, such as a count split for a median, closure of a census against the plotted total, or a third point on a fitted line, a closing check verifies it before the answer is stated. The final step states the graded answer exactly. Questions whose correct answer is an instructed refusal token, for example when an asked ageing mode is not among the plotted panels or when the cohort median never reaches the threshold inside the plotted window, receive traces that walk the presence or censoring check before concluding with the token.

Each question type is generated by at least one dedicated script, and most admit more than one genuinely different solution path. The library realises 56 template styles across the twelve families, and the styles differ in which reads they take and in what order, not merely in wording. For example, one quadrant-count style tallies all four quadrants and checks that they sum to the plotted total, while its sibling filters by sign along one axis first, and one fade-rate style works from the endpoints of the steepest curve while its sibling normalises the capacity drop before scaling. The style for each row is selected deterministically from the row identifier, so the library is exactly reproducible.

## A worked example

The trace below ships in the `formation_protocol_fade_band` config and is reproduced verbatim, together with the exact chart image from that row, so the reasoning can be followed against the pixels. It works a chain of materials-science concepts end to end: the cohort is a set of cells that received different formation protocols, the first-cycle conditioning step that sets a lithium-ion cell's interphase and downstream life, and each cell's end of life follows the per-cell convention of 80 per cent of its own initial capacity. The cohort's trajectories bunch together in the flat pre-knee band before fanning out at their knees, and because cohort sizes vary across the dataset and the larger cohorts are genuinely too dense to count by eye, the traces describe the cohort without asserting a specific cell count. It then calibrates the cycle axis, identifies the steepest-fading cell, reads its capacity endpoints off the logarithmic axis, normalises the capacity drop to the cell's own initial capacity, converts it into the standard fade-rate unit of per cent lost per thousand cycles, and then openly distinguishes that two-point estimate from the least-squares convention the graded answer follows.

![Formation-cell capacity-fade trajectories on a logarithmic capacity axis, coloured by knee-onset cycle, with a dashed end-of-life reference line](assets/worked_example.png)

Question: Among the plotted cells, identify the one whose capacity-fade trajectory has the steepest overall slope and estimate that cell's fade rate in percent of initial capacity lost per 1000 cycles.

```text
This figure shows the capacity-fade trajectories of a battery-formation
cohort of cells, plotted as discharge capacity in Ah on a logarithmic
scale against cycle number. A dashed reference line marks
the end-of-life capacity. Each cell's end of life is where its capacity
first drops to 80 per cent of its own initial value.

The cycle axis carries major ticks every 200 cycles. The capacity axis
is logarithmic, with decade ticks at 0.001, 0.01 and 0.1 Ah.

The steepest trajectory starts near 0.245 Ah at cycle 3.

The same trajectory ends near 0.011 Ah at cycle 950, a drop of 0.234 Ah
over 947 cycles.

Normalising the drop first: 0.234 Ah is 95.5 per cent of the cell's
initial 0.245 Ah, over 947 cycles, so the chord rate is roughly 100.8
per cent per 1,000 cycles. Refining with a least-squares fit through all
of the cell's plotted points gives 81.4 per cent per 1,000 cycles.

The steepest-fading cell loses capacity at 81.404 %/kcycle.
```

The final answer, 81.404 per cent per thousand cycles, is the exact pipeline value and grades correct under the parent benchmark's tolerances. The gap between the two-point chord estimate and the graded value is itself instructive: the trajectory stays flat for hundreds of cycles before its knee, so a fit through all of the plotted points reads shallower than the endpoint chord.

## Validation

Every trace passes three checks at build time, and a single failure aborts the build. The final answer must grade 1.0 against the parent benchmark row under the parent benchmark's max-of-three tolerance grader, the trace must contain no placeholder or not-a-number leak, and the step count must meet the orientation-reads-reasoning minimum. In addition, every generator that rebuilds data through the pipeline cross-checks its recomputed headline values against the stored instance ground truth before emitting any trace, so a drifted rebuild cannot silently produce a plausible but wrong intermediate. Samples of the built traces were also checked against their rendered charts before release.

## Schema

Each config is one chart family and each row is one trace. The flat columns support analysis and the `messages` column supports trainers that consume chat-format records directly.

| column | type | content |
|---|---|---|
| `question_uid` | string | joins to the parent benchmark row (`{instance_id}::{question_id}`) |
| `instance_id` | string | the chart instance the row belongs to |
| `chart_family` | string | one of the twelve family names |
| `question_id` | string | the question type within the family |
| `image` | image | the chart PNG, identical bytes to the parent benchmark row |
| `question` | string | the prompt, verbatim from the parent benchmark row |
| `reasoning` | string | the step-by-step trace as prose paragraphs |
| `answer` | string | the final answer with units, grading correct under the parent benchmark's tolerances |
| `messages` | list | a two-turn chat record: the user turn carries the question and the assistant turn carries the reasoning inside `<think>` tags followed by the answer |
| `trace_style` | string | which solution-path template produced the trace |
| `generator` | string | generator script and version, for provenance |
| `license` | string | CC-BY-NC-4.0 |

## How to load

The dataset is hosted at `translorentz/matsci-visual-reasoning-nc`. Each chart family is a separate config, and loading a config returns that family's trace rows.

```python
from datasets import load_dataset

ds = load_dataset(
    "translorentz/matsci-visual-reasoning-nc",
    "fade_trajectory_protocol_overlay",
    split="train",
)
row = ds[0]
print(row["question"])
print(row["reasoning"])
print(row["answer"])
```

## Use cases

The library serves three purposes. For training, it provides contamination-free chain-of-thought supervision for chart reading in a specialist scientific domain, at a scale where hand-authoring is impossible but programmatic generation stays exact. For evaluation, it lets graders score the path a model takes rather than only the endpoint, which makes error analysis concrete because each miss can be attributed to a specific read or reasoning step. For tool builders, each trace spells out precisely the intermediate quantities a pixel-reading or axis-calibration tool would need to supply, which turns the gap between vanilla vision models and the exact ceiling into a specification.

## Relationship to the parent benchmark

The images, questions, answers and grading tolerances are exactly those of MatSciChartQ at version 1.8.1-nc, whose July 2026 enrichment cut strengthened the materials-science framing of the questions and added six question types grounded in degradation-mode attribution, knee onset, electrode balance, and protocol dose-response reading. This repository adds only the reasoning traces and the chat-format packaging. Baseline pass rates for the parent benchmark, including the Blind floor that confirms the questions cannot be answered from text priors alone, are reported on the parent dataset card.

## Data sources and citation

The charts re-render numerics from four CC BY-NC 4.0 upstream datasets. Only the sources that are directly needed to generate the images are cited, and all four are listed below.

```bibtex
@article{vanvlijmen2025aging,
  title   = {Aging matrix visualizes complexity of battery aging across hundreds of
             cycling protocols},
  author  = {van Vlijmen, Bruis and others},
  journal = {Energy \& Environmental Science},
  volume  = {18},
  pages   = {6641--6654},
  year    = {2025},
  doi     = {10.1039/d4ee05609d}
}

@article{che2025diagnostic,
  title   = {Diagnostic-free onboard battery health assessment},
  author  = {Che, Yunhong and Lam, Vivek N. and Rhyu, Jinwook and Schaeffer, Joachim
             and Kim, Minsu and Bazant, Martin Z. and Chueh, William C. and
             Braatz, Richard D.},
  journal = {Joule},
  volume  = {9},
  number  = {8},
  pages   = {102010},
  year    = {2025},
  doi     = {10.1016/j.joule.2025.102010}
}

@article{cui2024formation,
  title   = {Data-driven analysis of battery formation reveals the role of electrode
             utilization in extending cycle life},
  author  = {Cui, Xiao and Kang, Stephen Dongmin and Wang, Sunny and others},
  journal = {Joule},
  volume  = {8},
  number  = {11},
  pages   = {3072--3087},
  year    = {2024},
  doi     = {10.1016/j.joule.2024.07.024}
}

@article{xie2023htpmd,
  title   = {A cloud platform for sharing and automated analysis of raw data from
             high throughput polymer MD simulations},
  author  = {Xie, Tian and Kwon, Ha-Kyung and Schweigert, Daniel and others},
  journal = {APL Machine Learning},
  volume  = {1},
  pages   = {046108},
  year    = {2023},
  doi     = {10.1063/5.0160937}
}
```

## Cite this dataset as

If you use these traces in your work, please cite the dataset itself as follows.

```bibtex
@misc{cheong2026matscichartqtraces,
  title     = {MatSciChartQ-Traces: grounded reasoning traces for
               materials-science chart question answering},
  author    = {Cheong, Bryan},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/9572},
  url       = {https://huggingface.co/datasets/translorentz/matsci-visual-reasoning-nc}
}
```

## Licence and provenance

The dataset is distributed under CC BY-NC 4.0, inherited from the upstream corpora. Build provenance: schema traces-1.0.0, generator version 1.4.1, built 2026-07-12 from parent deposit version 1.8.1-nc. Per-family trace counts and the validation statement are recorded in `manifest.json` at the repository root.