Download SPAN_ES/input_format.json from iLuvU2/figrefs-models: direct link, hf CLI and curl.
- Browser
- Download file 2.43 kB
-
https://huggingface.co/iLuvU2/figrefs-models/resolve/main/SPAN_ES/input_format.json
- Command line
-
hf download hf://iLuvU2/figrefs-models/SPAN_ES/input_format.json
-
curl -L -o input_format.json https://huggingface.co/iLuvU2/figrefs-models/resolve/main/SPAN_ES/input_format.json
2.43 kB
| { | |
| "version": "run4-v1", | |
| "config": "SPAN_ES", | |
| "stage": "SPAN", | |
| "max_len": 2048, | |
| "base": "answerdotai/ModernBERT-large", | |
| "input": "Run 4: span tagger with EARLY STOPPING, without (SPAN_ES) and with (SPAN_VIS) text from the figure image.\n\nOne script; each Kaggle account runs ONE config (2xT4):\n python run4_kaggle.py launch SPAN_ES | python run4_kaggle.py launch SPAN_VIS\nInputs: run1_data.jsonl; for SPAN_VIS also figtext.py and the figimg/ folder (cropped figures + index.json).\nOutput: /kaggle/working/run4_<cfg>/ and run4_<cfg>_scores.zip.\n\nModel = run 3's SPAN (ModernBERT-large span tagger: \"[CLS] Figure F. <caption> [SEP]\" + 2,048-token windows tiling\nthe whole paper, one [MASK] marker per sentence, A3 + paraphrase swap, pos_weight <= 8, batch 8, lr 2e-5).\nEarly stopping (both configs), deliberately conservative: validation = the 3 former test papers (17483019,\n19673075, 23041342; never trained on). Up to 8 epochs, linear decay over 8 epochs; every half epoch the model\nscores the validation papers; metric = average precision over (figure, sentence) pairs (threshold-free), smoothed\nover the last two checks. Never stops before epoch 3; stops only after 4 checks (2 epochs) without a new smoothed\nbest; then restores the best weights. Validation loss is logged too. The held-out fold of the 5-fold rotation is never looked\nat during training, so out-of-fold scores stay honest; validation papers are used only for stopping.\nSPAN_VIS: before training, Qwen2-VL-2B-Instruct (figtext.py, fp32, one shard per GPU) writes for every cropped\nfigure (a) the text inside it (OCR) and (b) a literal visual description; the header becomes\n\"Figure F. <caption> Figure text: <ocr> Figure shows: <description>\" (up to 400 tokens instead of 200). Figures with\nno crop get \"Figure shows: (no image available)\".\nProtocol: 5 folds x seeds 1,2,3 on the 15 training papers -> oof/<cfg>.json; final model on all 15 (early-stopped\non the validation papers) -> models/<cfg>/ and val/<cfg>.json (its scores on the validation papers).\nRobust: failed jobs are logged and skipped; re-running launch resumes.\nEnv: R1_SEEDS R1_OUT; R1_SMOKE=1 quick check (tiny models, 2 folds, 1 seed, 2 epochs, 40 sentences);\n R1_SMOKE_REAL=1 with real models; R1_FIGTEXT_LIMIT=n caps figures for figtext in smoke tests.", | |
| "decision": "per-article z-score of the logit >= threshold (threshold from phase8.py)" | |
| } |