| # Data archive |
|
|
| ## Training data |
|
|
| `training/` contains only train-ready data referenced by a released portable |
| configuration: |
|
|
| - `data_joint_tsqa/`: shared Joint-GRPO pool used by B0, R2, and R3. |
| - `ablation/caption_system_prompt_ablation_v1/`: corrected Joint-SFT pairs for |
| R1/B0. |
| - `ablation/d1_ours_*`: internal-data-only SFT and GRPO inputs. |
| - `ablation/d2_tsqa_*`: TSQA-only SFT and GRPO inputs. |
| - `ablation/d3_gpt5nano_final_alignment/`: validated GPT-perceived SFT and GRPO |
| targets. API request receipts and intermediate batches are intentionally not |
| included. |
| - `ablation/r3_no_easyfirst/` and `data_{univar,bivar,multivar,replay}/`: |
| static epoch manifests and raw series for the complete R3 upstream path. |
| - `ablation/b0_joint_sft_caption_prompt_v2/`: corrected final SFT manifests |
| used after the R3 upstream path. |
|
|
| The duplicate `ablation_V0` archive and run logs, locks, PID files, caches, and |
| failed diagnostic outputs are not part of the release. |
|
|
| ## Evaluation data |
|
|
| - `evaluation/metric_qa/requests_paper_um2000.jsonl`: paper U+M subset. |
| - `evaluation/caption/*_paper_um2000.jsonl`: paper U+M request and hidden-GT |
| subsets. |
| - `evaluation/tsqa/series_overlap_audit.json`: 255-request exclusion audit. |
| - `evaluation/tsqa/requests_clean3264.jsonl`: deterministic clean subset used |
| directly for inference and scoring. |
| - `evaluation/PROVENANCE.json`: omitted 3,000-row source hashes and the exact |
| deterministic filtering rules used to derive the released 2,000-row files. |
|
|
| Hashes, sizes, and JSONL row counts are recorded in |
| `manifests/data_manifest.json`. |
|
|
