Repository Public
1{2 "id": "flan-t5-large-htr",3 "name": "Flan-T5-Large-HTR",4 "family": "hf-seq2seq",5 "version": "0.1.0",6 "n_params": 783150080,7 "base_model": "google/flan-t5-large",8 "licence": "Apache-2.0 base model (Google); fine-tuned weights for non-commercial research use (training pairs include CC BY-NC sources)",9 "default": false,10 "speed_hint": "≈1.5 s per paragraph, ≈6 s for 3 000 characters (greedy, Apple GPU)",11 "max_len": 512,12 "max_new_tokens": 512,13 "metrics": {14 "reference": "Differences in chrF++ versus the identity baseline (B0: return the input unchanged) on the Medium Haiku rewrite→original pairs; test = document-safe test split, held-out domain/period = generalisation splits of the paper. Exploratory unless marked confirmatory. Number recall and placeholder exactness are preservation rates on the test split; Pangram = share of reconstructions the Pangram classifier labels Human (54 passages).",15 "val_chrf_delta": 2.04,16 "val_chrf_pm": 0.15,17 "test_chrf_delta": 1.79,18 "test_chrf_pm": 0.07,19 "test_confirmatory": false,20 "heldout_domain_chrf_delta": 1.27,21 "heldout_domain_pm": 0.18,22 "heldout_period_chrf_delta": 1.09,23 "heldout_period_pm": 0.14,24 "number_recall": 0.986,25 "placeholders_exact": 0.992,26 "pangram_human": 54,27 "pangram_n": 54,28 "pangram_mixed": 0,29 "pangram": {30 "detector": "Pangram 4.1",31 "date": "2026-11-10",32 "n": 54,33 "human_share": 1.0,34 "flagged_share": 0.0,35 "mean_ai_likelihood": 0.0,36 "mean_ai_likelihood_ci": [37 0.0,38 0.039 ],40 "labels": {41 "human": 54,42 "mixed": 0,43 "ai": 044 },45 "reference": {46 "originals": {47 "human_share": 1.0,48 "flagged_share": 0.0,49 "mean_ai_likelihood": 0.050 },51 "rewrites": {52 "human_share": 0.537,53 "flagged_share": 0.463,54 "mean_ai_likelihood": 0.187555 }56 },57 "source": "artifacts/metrics/pangram_test_htr2/pangram_summary.md (system B2_flan-t5-large; flagged = labelled Mixed or AI; mean = word-weighted ai_assistance_score, 95 % document-clustered bootstrap CI)",58 "note": "This measures one commercial detector's behaviour on a 54-passage sample. A low AI-likelihood is not evidence of human authorship, and HTR must not be used to evade AI-text detection or for integrity decisions."59 }60 },61 "notes": [62 "Pretrained backbone: its pretraining corpus may contain pre-2018 scientific text; memorisation controls cover only the HTR training pairs (see the paper's threats-to-validity section).",63 "Reads at most 512 tokens (about 2 000 characters): longer inputs are truncated and a warning is returned.",64 "System B2 of the paper (fine-tuned pretrained encoder–decoder baseline)."65 ],66 "trained_on": "Fine-tuned on the same 72 698 validated Haiku rewrite → original pairs (Medium stage of WP-AI-0002, same document-safe splits), context given as a text prefix; placeholders spelled [[C1]] for the pretrained tokenizer.",67 "exported_at": "2026-11-10",68 "hf_repo": "spboucher04/flan-t5-large-htr",69 "hf_url": "https://huggingface.co/spboucher04/flan-t5-large-htr"70}