Repository Public
1{2 "id": "qwen2.5-3b-htr",3 "name": "Qwen2.5-3B-HTR",4 "family": "hf-causal",5 "version": "0.1.0",6 "n_params": 3085938688,7 "base_model": "Qwen/Qwen2.5-3B",8 "licence": "Apache-2.0 base model (Alibaba Qwen); fine-tuned weights for non-commercial research use (training pairs include CC BY-NC sources)",9 "default": false,10 "speed_hint": "≈3–5 s per paragraph (greedy, bf16, Apple GPU)",11 "max_new_tokens": 768,12 "metrics": {13 "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).",14 "val_chrf_delta": 2.42,15 "val_chrf_pm": 0.13,16 "test_chrf_delta": 2.27,17 "test_chrf_pm": 0.06,18 "test_confirmatory": false,19 "heldout_domain_chrf_delta": 2.05,20 "heldout_domain_pm": 0.13,21 "heldout_period_chrf_delta": 1.91,22 "heldout_period_pm": 0.14,23 "number_recall": 0.998,24 "placeholders_exact": 0.999,25 "pangram_human": 49,26 "pangram_n": 54,27 "pangram_mixed": 5,28 "pangram": {29 "detector": "Pangram 4.1",30 "date": "2026-11-10",31 "n": 54,32 "human_share": 0.907,33 "flagged_share": 0.093,34 "mean_ai_likelihood": 0.0257,35 "mean_ai_likelihood_ci": [36 0.0053,37 0.052338 ],39 "labels": {40 "human": 49,41 "mixed": 5,42 "ai": 043 },44 "reference": {45 "originals": {46 "human_share": 1.0,47 "flagged_share": 0.0,48 "mean_ai_likelihood": 0.049 },50 "rewrites": {51 "human_share": 0.537,52 "flagged_share": 0.463,53 "mean_ai_likelihood": 0.187554 }55 },56 "source": "artifacts/metrics/pangram_test_qwen3b/pangram_summary.md (system LLM_qwen2.5-3b; flagged = labelled Mixed or AI; mean = word-weighted ai_assistance_score, 95 % document-clustered bootstrap CI)",57 "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."58 }59 },60 "notes": [61 "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).",62 "Highest number and placeholder fidelity of all systems (99.8 % / 99.9 % on the test split); second-best lexical gain after Flan-T5-XL (exploratory system LLM_qwen2.5-3b of the paper).",63 "Decoder-only model: the rewritten passage is given in a prompt and the reconstruction is generated after it; greedy decoding only."64 ],65 "trained_on": "Supervised fine-tuning on the same 72 698 validated Haiku rewrite → original pairs (Medium stage of WP-AI-0002, same document-safe splits; loss on the original only), 6 000 steps of effective batch 16 on one H100, best validation CE 0.701; placeholders spelled [[C1]]. Weights stored in bfloat16.",66 "exported_at": "2026-11-11",67 "hf_repo": "spboucher04/qwen2.5-3b-htr",68 "hf_url": "https://huggingface.co/spboucher04/qwen2.5-3b-htr"69}