Repository Public
1{2 "id": "htr-base-0.1.0",3 "name": "HTR-Base 0.1.0",4 "family": "htr",5 "arch": "m1",6 "version": "0.1.0",7 "n_params": 334844929,8 "base_model": null,9 "licence": "Apache-2.0 (code and weights); non-commercial research use where CC BY-NC sources were included — see LICENSE-MODEL.txt",10 "default": false,11 "speed_hint": "≈0.7 s greedy, ≈1.1 s beam 3 per paragraph (Apple GPU)",12 "max_len": 640,13 "text_meta_prefix": false,14 "metrics": {15 "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).",16 "val_chrf_delta": 0.83,17 "test_chrf_delta": 0.69,18 "test_chrf_pm": 0.05,19 "test_confirmatory": false,20 "heldout_domain_chrf_delta": 0.29,21 "heldout_period_chrf_delta": -0.12,22 "number_recall": 0.992,23 "placeholders_exact": 0.993,24 "pangram_human": 53,25 "pangram_n": 54,26 "pangram_mixed": 0,27 "pangram": {28 "detector": "Pangram 4.1",29 "date": "2026-11-10",30 "n": 54,31 "human_share": 0.9815,32 "flagged_share": 0.0185,33 "mean_ai_likelihood": 0.0025,34 "mean_ai_likelihood_ci": [35 0.0,36 0.007237 ],38 "labels": {39 "human": 53,40 "mixed": 1,41 "ai": 042 },43 "reference": {44 "originals": {45 "human_share": 1.0,46 "flagged_share": 0.0,47 "mean_ai_likelihood": 0.048 },49 "rewrites": {50 "human_share": 0.537,51 "flagged_share": 0.463,52 "mean_ai_likelihood": 0.187553 }54 },55 "source": "artifacts/metrics/pangram_test_htr2/pangram_summary.md (system htr-m1-base-lambda; flagged = labelled Mixed or AI; mean = word-weighted ai_assistance_score, 95 % document-clustered bootstrap CI)",56 "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."57 }58 },59 "notes": [60 "Larger HTR Transformer (d_model 1024, 8+8 layers) trained with the same recipe as HTR-Tiny; exploratory results.",61 "Architecture m1 does not condition on the context fields."62 ],63 "trained_on": "Medium stage of WP-AI-0002: denoising pretraining on 258 264 rights-verified pre-2018 passages (PMC Open Access, CC BY economics and finance journals, arXiv, public-domain books), then fine-tuning on 72 698 validated Haiku rewrite → original pairs of 14 000 passages (32k SentencePiece).",64 "exported_at": "2026-11-10T22:20:40Z",65 "hf_repo": "spboucher04/htr-base",66 "hf_url": "https://huggingface.co/spboucher04/htr-base"67}