|
..
|
|
audit_run.py
|
RQ6 audits (memorisation, agreement, ambiguity) on Medium validation outputs |
|
|
batch_admin.py
|
Eval suite runner, blind human-evaluation sheet generator |
|
|
batch_wait.py
|
Medium Haiku job submitted (120 000 requests); batch_wait helper; STATUS |
|
|
build_corpus.py
|
WP-AI-0002: repository scaffold, licence governance, PMC S3 ingest, JATS parser, segmentation,... |
|
|
build_figures.py
|
Figures: split label |
|
|
build_medium_corpus.sh
|
build_medium_corpus: configurable fixed splits |
|
|
decode_search.py
|
B2 backbone: bf16/cosine/patience in hf_finetune, multi-split predictions, bart-large chain;... |
|
|
discover_arxiv.py
|
Add arXiv/LaTeX, Crossref/PDF and books adapters (agents), governance docs, KV-cache decoding,... |
|
|
discover_books.py
|
Add arXiv/LaTeX, Crossref/PDF and books adapters (agents), governance docs, KV-cache decoding,... |
|
|
discover_crossref.py
|
Eval suite runner, blind human-evaluation sheet generator |
|
|
discover_pmc.py
|
WP-AI-0002: repository scaffold, licence governance, PMC S3 ingest, JATS parser, segmentation,... |
|
|
evaluate.py
|
Configurable bootstrap count, confirmatory flag, freeze script |
|
|
export_release.py
|
Add arXiv/LaTeX, Crossref/PDF and books adapters (agents), governance docs, KV-cache decoding,... |
|
|
freeze_preregistration.py
|
Eval suite runner, blind human-evaluation sheet generator |
|
|
h4_w1_bootstrap.py
|
Confirmatory scoring (6 sets, 10 000 resamples), H4 bootstrap, confirmatory... |
|
|
lambda_chain_large.sh
|
lambda: Large chain with 8192/6144-token batches (12288 OOMs at 71 GB) |
|
|
lambda_chain_t5large.sh
|
MRT (minimum-risk training) module + script; t5-large chain batch 12 / max-len 448 |
|
|
lambda_setup.sh
|
Pangram detector-sensitivity evaluation (RQ7): client, script, docs, results; Lambda GPU... |
|
|
lambda_sync.sh
|
Pangram detector-sensitivity evaluation (RQ7): client, script, docs, results; Lambda GPU... |
|
|
lambda_train_chain.sh
|
training: --num-workers for DataLoader (CPU noising/tokenisation starved the H100 at 0 % util) |
|
|
lambda_train_chain2.sh
|
lambda: second-instance chain (Large on GPU 0, Base seed 2 on GPU 1) |
|
|
make_human_eval_sheet.py
|
Eval suite runner, blind human-evaluation sheet generator |
|
|
make_splits.py
|
Add arXiv/LaTeX, Crossref/PDF and books adapters (agents), governance docs, KV-cache decoding,... |
|
|
medium_rewrite_job.sh
|
D20: Medium-stage rewrites with Haiku 5.5 (author-approved); medium job script |
|
|
pangram_eval.py
|
Pangram detector-sensitivity evaluation (RQ7): client, script, docs, results; Lambda GPU... |
|
|
pilot_rewrite_job.sh
|
Add arXiv/LaTeX, Crossref/PDF and books adapters (agents), governance docs, KV-cache decoding,... |
|
|
predict_b2_hf.py
|
B2 T5-small baseline results (val, exploratory), predict_b2_hf script |
|
|
predict_confirmatory.sh
|
Medium val: M3-light, M3 token-only seed 2; confirmatory prediction script |
|
|
predict_heldout_pilot.sh
|
Confirmatory plan, held-out rewriter prediction script, STATUS seed 3 |
|
|
predict_tagger.py
|
STATUS: cluster usage |
|
|
predict.py
|
Batched greedy reconstruction (identical outputs, faster); predict.py --batch-size |
|
|
pretrain_denoise.py
|
training: --num-workers for DataLoader (CPU noising/tokenisation starved the H100 at 0 % util) |
|
|
relabel_disciplines.py
|
Eval suite runner, blind human-evaluation sheet generator |
|
|
rewrite.py
|
Add arXiv/LaTeX, Crossref/PDF and books adapters (agents), governance docs, KV-cache decoding,... |
|
|
run_confirmatory.sh
|
M* exported as HTR-Tiny 0.3.0 and deployed to htr.spboucher.ai; confirmatory runner script |
|
|
run_eval_suite.sh
|
Eval suite runner, blind human-evaluation sheet generator |
|
|
train_b2_hf.py
|
MRT (minimum-risk training) module + script; t5-large chain batch 12 / max-len 448 |
|
|
train_discriminator.py
|
backbone: flan-t5-large instead of bart-large (NaN gradients with torch 2.11) |
|
|
train_mrt.py
|
train_mrt: use ReconstructionResult.reconstruction |
|
|
train_tagger.py
|
STATUS: cluster usage |
|
|
train_tokenizer.py
|
Medium corpus v0 (13 372 docs / 371 419 eligible passages) summaries; tokenizer trainer on... |
|
|
train.py
|
training: --num-workers for DataLoader (CPU noising/tokenisation starved the H100 at 0 % util) |
|
|
validate_pairs.py
|
WP-AI-0002: repository scaffold, licence governance, PMC S3 ingest, JATS parser, segmentation,... |
|
|
vertical_slice.sh
|
Add arXiv/LaTeX, Crossref/PDF and books adapters (agents), governance docs, KV-cache decoding,... |
|