# ThrowGen: datasets and results

This archive contains the evaluation datasets, model predictions, and
evaluation results behind the numbers, tables, and figures reported in the
ThrowGen paper. It holds the artifacts the reported results are computed from,
and nothing else.

```
published_data/
  data/throwgen/<dataset>/         # evaluation and validation datasets
  results/llm_output/<dataset>/    # raw LLM predictions
  results/metrics/<dataset>/       # evaluation metrics
  results/qualitative/             # manual qualitative analysis
  MANIFEST.txt                     # every file in this archive, one per line
```

## Datasets (`data/throwgen/`)

Six of the seven datasets trace the construction funnel described in the
dataset section of the paper; the seventh is the validation set used for prompt
selection.

| Dataset | Methods | Role |
|---|---|---|
| `real-non-direct-mega-test-data` | 1,099 | Initial collection from GitHub projects |
| `real-mega-test-data` | 546 | After keeping exceptions thrown directly in the target method |
| `real-mega-test-data-with-exception` | 525 | After requiring the exception type to match the test |
| `real-mega-test-data-with-exception-with-project-with-gold` | 399 | After requiring the developer implementation to pass its own tests |
| `real-mega-test-data-with-exception-with-project-with-gold-with-throw` | 304 | **Evaluation set** — all results below are on this set |
| `mega-val-data-with-exception-with-project-with-gold-with-throw` | 214 | Validation set |
| `all-data` | 518 | Evaluation and validation sets combined |

Each dataset is stored as one JSON Lines file per field, aligned row by row:
`mut` (the target method), `mut_no_throw` (the method with exception-raising
code removed, which is the task input), `ebts` (exception-behavior tests),
`nebts` (non-exceptional tests), `randoop_tests` and `evosuite_tests`
(tool-generated tests), and the contextual fields `class_info`, `method_info`,
`throw_info`, `import_info`, `local_variable_type`, `coverage`,
`thrown_exception`, and `unreported_exception`. `id.jsonl`, `project.jsonl`,
`start_line.jsonl`, and `end_line.jsonl` locate each method in its source
project. `dataset_stats.json` and `throw-count.json` carry the aggregate
statistics reported in the dataset tables.

## Models and prompt configurations

Five models, all evaluated on the 304-method evaluation set, 10 samples per
method at temperature 0.8:

- `llama3.1:8b-instruct-q8_0` (`llama_cpp`)
- `phi4:14b-q8_0` (`llama_cpp`)
- `qwen2.5-coder:7b-instruct-q8_0` (`llama_cpp`)
- `qwen2.5-coder:32b-instruct-q8_0` (`llama_cpp`) — the main model
- `gpt-5-mini` (`azure`)

Prompt configurations:

- `base` and `tuctn-all-info` for all five models, for the model comparison
- `cmtu`, `only-avsym`, `only-lcov`, `only-nebt`, and `only-threxc` for the
  main model, for the prompt comparison and context ablations
- `base-repair@1..4` and `tuctn-all-info-repair@1..4` for the main model, for
  the iterative self-repair experiment

## Results

`results/llm_output/` — one JSON Lines file per model and prompt
configuration, holding the raw generations for each method.

`results/metrics/` — for each model and prompt configuration, a
`-summary.json` (aggregate) and a `-each-sample.jsonl` (per method) for the
three evaluation types the paper reports:

- `run-ebts-pass-at-k` — compile@k and pass@k on the developer-written
  exception-behavior tests
- `run-all-pass-at-k` — pass@k on those tests together with the non-exceptional
  tests
- `run-all-with-tools-pass-at-k` — the above plus the Randoop and EvoSuite
  generated tests

Two additional per-sample files,
`run-all-llama_cpp-qwen2.5-coder:32b-instruct-q8_0-base-multi_ebt-each-sample.jsonl`
and its `tuctn-all-info` counterpart, record which individual methods pass
under each prompt; they are the source of the Venn diagram and the overlap
counts in the results section.

File names follow
`<evaluation-type>-<backend>-<model>-<prompt>-<setup>-{summary.json,each-sample.jsonl}`,
where `setup` is `multi_ebt` for the main experiments and `repair` for the
self-repair iterations.

`results/qualitative/` — the manual analysis behind the qualitative section:
per-sample labels and summaries for `base` and `tuctn-all-info`, and the
inter-annotator agreement statistics. Files prefixed `same_samples_` restrict
the analysis to the methods labeled under both prompt configurations.

## Reproducing the reported numbers

Every aggregate metric in this archive was checked against the corresponding
value printed in the paper:

- `run-ebts-pass-at-k` — all 138 values match.
- `run-all-pass-at-k` — all 138 values match.
- `run-all-with-tools-pass-at-k` — 30 of 90 values match. The four `only-*`
  context ablations agree exactly; the remaining eleven model and prompt
  combinations differ by 0.46 percentage points on average and 1.95 at worst
  (`qwen2.5-coder:32b-instruct-q8_0` with `tuctn-all-info`, compile@1). The
  tool-augmented evaluation was re-run after those numbers were typeset, so the
  values here supersede the ones printed in the paper.

One column is not reproducible from this archive: the tool-augmented pass@k
figures for the self-repair iterations. The per-iteration
`run-all-with-tools-pass-at-k` summaries for the `repair` setup are unavailable,
so only the `run-ebts-pass-at-k` and `run-all-pass-at-k` columns of the
self-repair table can be regenerated here.