jackmikea/barbados-htr-final-submission

R.O.A.D. Barbados Historic Handwriting — trained checkpoints (final submission)

Trained weights of the 25 ensemble members of our final Zindi submission (line-level HTR of 17th–18th century Barbados legal manuscripts). They are downloaded by download_weights.sh of the solution package into outputs/runs/ and outputs/runs_htr/; run_inference.sh re-creates the submission from them.

path	what	base model (downloaded separately from its original repo)
runs/f2_9b_lora64_ls01_s{42,7,123,99,2024}/checkpoint-3843	LoRA r64 (5 seeds)	Qwen/Qwen3.5-9B
runs/f1_9b_h128_aug4_lora64_s42/checkpoint-3843	LoRA r64	Qwen/Qwen3.5-9B
runs/f1_9b_h128_aug4_s42/soup	full fine-tune, average of the last 2 epochs (complete model)	–
runs/f2_4b_lora64_ls01_s{42,7}/checkpoint-3843, runs/f_4b_h144_full/checkpoint-771	LoRA r64	Qwen/Qwen3.5-4B
runs/f3_9bbase_lora64_ls01_s42/checkpoint-3843	LoRA r64	Qwen/Qwen3.5-9B-Base
runs/f3_27b_lora64_ls01_s42/checkpoint-3843	LoRA r64	Qwen/Qwen3.5-27B
runs/f_9b_h96_full/checkpoint-2307	LoRA r64	Qwen/Qwen3.5-9B
runs/f3_q3vl8b_lora64_ls01_s42/checkpoint-2562, runs/f_q3v_{ls02,aug14,h144,h112,h144_ls02}_full/checkpoint-*	LoRA r64	Qwen/Qwen3-VL-8B-Instruct
runs/f_q3vl4b_h144_full/checkpoint-771	LoRA r64	Qwen/Qwen3-VL-4B-Instruct
runs_htr/h_px_base_full/best	Pix2Struct fine-tune (complete model)	–
runs_htr/h_court_full/best, runs_htr/h_trocrlarge_full/best	TrOCR fine-tunes (complete models)	–
runs/cf2_ctc_big_aug14{,_s7}/best.pt (+ meta.json, syms.json)	PyLaia CRNN-CTC from scratch	–

adapter_config.json files keep the base-model path used at training time; download_weights.sh rewrites it to the local models/<name> directory. Every runs/<job>/job.json is the exact training configuration (see configs/jobs.yaml).