File size: 5,719 Bytes
90d6c2f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 | # ChartBench-E (Rebuilt, Insight-Oriented)
This implementation rebuilds ChartBench-E and extends it into a more diagnostic benchmark for VLM chart reading.
## What it provides
- 20 data patterns grouped into 6 behavior families.
- Multiple chart morphologies: `line`, `bar_v`, `bar_h`, `stacked`.
- Annotation modes: `labeled`, `partial`, `unlabeled`.
- Multi-factor difficulty beyond point count:
- visual perturbations (noise, blur, JPEG artifacts, partial occlusion)
- semantic traps (truncated axis, log-scale reading, legend confusion)
- event-level stressors (outlier spikes, near-tie values, regime shifts)
- Diagnostic metadata tags for capability slicing:
- `event_tags`
- `perceptual_tags`
- `semantic_risk_tags`
- Reproducible generation with random seed.
- Evaluation metrics:
- SPR (Success Parsing Rate)
- OA (Overall Accuracy)
- NRA (Numerical Reading Accuracy)
- ARE (Average Relative Error)
## Install
```bash
cd chartbench_e
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
```
## Build benchmark-focused dataset
```bash
PYTHONPATH=. python scripts/build_chartbench_e.py \
--out output/chartbench_e_benchmark_plus \
--total-samples 20000 \
--seed 42 \
--task-mode benchmark \
--challenge-level standard
```
Challenge levels:
- `standard`: light perturbation, mostly clean charts.
- `advanced`: balanced stress setting.
- `extreme`: heavy perturbation and denser hard cases.
## Make fixed evaluation subset
```bash
PYTHONPATH=. python scripts/make_eval_subset.py \
--index output/chartbench_e_benchmark_plus/metadata/index.jsonl \
--out output/chartbench_e_benchmark_plus/metadata/eval_subset.jsonl \
--per-task 400
```
Risk-focused subset example:
```bash
PYTHONPATH=. python scripts/make_eval_subset.py \
--index output/chartbench_e_benchmark_plus/metadata/index.jsonl \
--out output/chartbench_e_benchmark_plus/metadata/eval_compositional.jsonl \
--per-task 200 \
--require-risk-tag compositional_reasoning
```
## Evaluate predictions
Prediction format (`jsonl`):
```json
{"sample_id": "cb_0000001", "values": [[...], [...]]}
```
Run:
```bash
PYTHONPATH=. python scripts/evaluate_predictions.py \
--truth output/chartbench_e_benchmark_plus/metadata/eval_subset.jsonl \
--pred your_predictions.jsonl \
--out output/report.json \
--tolerance 0.05
```
## Multi-model benchmark (new)
### Supported model registry
See `scripts/model_registry.json`.
Default enabled models:
- `gpt-5.2`
- `gpt-5.2-high`
- `gpt-5`
- `gpt-5-mini-high`
- `qwen3-vl-30b-a3b-thinking`
- `qwen2.5-vl-72b`
- `glm-4.6v`
Explicitly excluded in this round:
- `deepseek-vl2`
- `paddleocr-vl-1.5`
### Required env vars
OpenAI provider:
- `OPENAI_API_KEY`
- `OPENAI_BASE_URL` (optional)
SiliconFlow provider:
- `SILICONFLOW_API_KEY`
- `SILICONFLOW_BASE_URL` (optional, default `https://api.siliconflow.cn/v1`)
Fallback rule for SiliconFlow models:
- if `SILICONFLOW_*` is missing, script falls back to `OPENAI_*`
### Prompt behavior
- The benchmark runner defaults to `--prompt-mode type_specific`.
- In `type_specific` mode, prompts are loaded from `prompts/chart_types/<chart_type>.txt`.
- Each prompt instantiation now includes a shape-aware JSON example through `{json_example}`, in addition to the strict JSON output contract.
- `prompts/chart_reading_prompt.txt` remains available for `--prompt-mode unified`.
### Run benchmark
Single model + 3 repeats + evaluation:
```bash
PYTHONPATH=. python scripts/run_vlm_benchmark.py \
--truth-index output/chartbench_e_benchmark_plus/metadata/eval_subset.jsonl \
--dataset-root output/chartbench_e_benchmark_plus \
--models gpt-5.2 \
--repeats 3 \
--max-workers 2 \
--max-retries 2 \
--timeout-seconds 120 \
--run-eval \
--tolerance 0.05
```
All enabled models:
```bash
PYTHONPATH=. python scripts/run_vlm_benchmark.py \
--truth-index output/chartbench_e_benchmark_plus/metadata/eval_subset.jsonl \
--dataset-root output/chartbench_e_benchmark_plus \
--models all \
--repeats 3 \
--max-workers 4 \
--max-retries 2 \
--timeout-seconds 120 \
--run-eval
```
Offline smoke test (no API call, sanity-only):
```bash
PYTHONPATH=. python scripts/run_vlm_benchmark.py \
--truth-index output/chartbench_e_benchmark_plus/metadata/eval_subset.jsonl \
--dataset-root output/chartbench_e_benchmark_plus \
--models gpt-5.2 \
--repeats 3 \
--max-samples 20 \
--run-eval \
--mock-from-truth
```
### Prompt A/B pilot
This pilot samples a 54-chart subset (18 chart types x 3 difficulty tiers x 1 sample), runs baseline vs. updated chart-type prompts, and writes a comparison report:
```bash
PYTHONPATH=. python scripts/run_prompt_ab_pilot.py \
--mock-from-truth
```
For real model runs, set provider credentials first and rerun without `--mock-from-truth`.
### Output structure
`output/bench_runs/{run_id}/`
- `run_config.json`
- `truth_subset.jsonl`
- `run_summary.json`
- `leaderboard.csv`
- `models/{model_key}/repeat_0x/pred.jsonl`
- `models/{model_key}/repeat_0x/report.json`
- `models/{model_key}/repeat_0x/failures.jsonl`
- `models/{model_key}/repeat_0x/raw_responses.jsonl`
- `models/{model_key}/repeat_0x/slice_report.json`
- `models/{model_key}/model_stability.json`
- `models/{model_key}/sample_stability.jsonl`
- `models/{model_key}/slice_summary.json`
Slicing coverage:
- `difficulty_tier` (L1-L4)
- `theme_domain`
- `axis_scale`
- `y_axis_truncated`
## Output structure (dataset)
- `images/*.png`: generated chart images.
- `metadata/index.jsonl`: full metadata with ground truth and challenge tags.
- `metadata/summary.json`: split and distribution summary.
- `metadata/eval_subset.jsonl`: fixed benchmark subset.
|