Reorganize train/eval configs by model directory and make total batch explicit

[iamtatsuki05]

WHY

The training configs under config/constract_llm/train/{pt,ft} were flat files, and the eval configs mixed several naming schemes (canonical case vs lowercase, abbreviated JP aliases such as sentence-llama-bi-1b, <Model>_<task>.json prefixes, model--task filenames), which made it hard to see which configs belong to the same model. In addition, the 1B/3B configs implicitly used size-scaled total batches (8192/16384 for pre-training, 1024/2048 for stage-2/fine-tuning at 8 GPUs) while the 0.5B recipe uses 4096/512, but nothing in the names said so, and the batch-matched 1B/3B recipe used by the current re-measurement campaigns had no canonical config in the repository at all.

WHAT

Train config layout

• train/pt/<Model>/{stage1,stage2}[-bs<total>].json and train/ft/<Model>/{pt,ft}[-bs<total>].json: one directory per model, stages as files (e.g. pt/Llama-EN-1B/stage1.json, ft/Sentence-Llama-Bi-EN-1B/ft-bs1024.json)
• Unsuffixed 1B/3B configs are new and match the 0.5B effective total batch on 8 GPUs (stage1 / WSL 4096, stage2 / FT 512); per-device batch is kept and gradient accumulation recomputed (3B FT halves per-device batch to fit 512)
• The former size-scaled 1B/3B configs carry an explicit -bs<total> suffix in file names, run_name, output_dir, hub ids and every eval reference (1B 8192/1024, 3B 16384/2048); 0.5B stays unsuffixed
• model/save_custom_model/llama_bi_*: unsuffixed configs for the batch-matched runs, *_bs<total>.json for the size-scaled ones (Llama-Bi stage-2 hub ids are suffixed the same way as the ModernBERT ones)

Eval config layout

• eval/pt/{glue,qa,isotropic} for pre-training models (suite names follow scripts/constract_llm/eval/pt/run_*.py; qa groups squad/csqa) and eval/ft/{mteb,bench_sbert,isotropic} for sentence models; mteb_full and mteb_full_chunks are folded into mteb/<Model>/{subset,full}.json + full_chunks/<task>.json
• Every suite uses one directory per model named after model_name_or_path, with the batch variant in the file name: <task>[-bs<total>].json for multi-config suites, default.json / bs<total>.json for single-config suites
• External baselines (sarashina2.2 base and its Bi conversion) and same-model per-language variants keep their project labels
• Matching eval configs are added for the new batch-matched 1B/3B models

Formatting

• learning_rate / adam_epsilon / weight_decay use the majority one-digit exponent notation (3e-4, 1e-5) instead of mixed 0.0003 / 3e-05
• 2-space indentation, trailing newline, and no redundant ./ prefix on path values
• README configuration lists (EN/JA) updated to the new paths

VERIFICATION

• All 591 config files parse; the pre-existing configs are byte-equivalent to the base after normalizing the -bs rename (no hyperparameter changes), and every new unsuffixed 1B/3B config satisfies per_device × 8 × grad_accum == 0.5B total
• Every eval file's directory equals the base name of its model_name_or_path and its file-name batch suffix equals the model's (checked for all iamtatsuki05/* references); task file names match task_name
• No remaining references to old paths in .py / .sh / .md / .yaml / .json; every config path mentioned in the READMEs exists; no code, test, or CI job globs the config directories
• Two read-only reviews (no blocking findings)
• Not covered here: the HF Hub repositories and local artifacts still use the pre-suffix names, so runs that consume existing -bs artifacts need the old ids until they are re-uploaded or renamed; the campaign-side overrides (eval_steps 400/800, save_total_limit 80/unlimited, use_reentrant: false for ST) are intentionally not folded into these configs