chore(format): apply prettier and ruff formatting

Author: dutifulbob

.agent/skills/general-visualization/SKILL.md

@@ -65,18 +65,18 @@ Prefer aligned plain text blocks over Markdown tables unless the user explicitly
 
 ### Side-by-Side Comparison (default, terminal-safe)
 
-Metric          fastText-ftz  transformer-int8  Change
-Scam precision  .92           .95 *             +3%
-Scam recall     .70           .78 ***           +8%
-Scam FPR        .018          .012              -1%
-Macro F1        .83           .85 *             +2%
+Metric fastText-ftz transformer-int8 Change
+Scam precision .92 .95 \* +3%
+Scam recall .70 .78 _\*\* +8%
+Scam FPR .018 .012 -1%
+Macro F1 .83 .85 _ +2%
 
 ### Comparison with Delta
 
-Metric          baseline-v1   candidate-v2      Change
-Model size MB   3.4           3.1 **            -9%
-Scam recall     .78 *         .76               -2%
-Latency p95 ms  58 **         61                +5%
+Metric baseline-v1 candidate-v2 Change
+Model size MB 3.4 3.1 ** -9%
+Scam recall .78 \* .76 -2%
+Latency p95 ms 58 ** 61 +5%
 
 ## Style Controls
 

.agent/skills/huggingface/SKILL.md

@@ -93,11 +93,13 @@ python3 scripts/sync_datasets_to_experiments_repo.py \
 ```
 
 Defaults:
+
 - x-posts dataset source: `data/sample.jsonl`
 - x-replies dataset source: `data/replies.jsonl`
 - snapshot id: auto-generated `yyyy-mm-dd-<petname>`
 
 Behavior:
+
 - Writes snapshot under `datasets/snapshots/<snapshot_id>/`.
 - Updates `datasets/INDEX.json`.
 - If content is unchanged (same hashes), does not create a duplicate snapshot.

config/thresholds.transformer.json

@@ -44,4 +44,4 @@
     }
   },
   "version": 1
-}
+}

docs/logs/2026-02-13.md

@@ -46,6 +46,7 @@ Recreated scripts/.venv with uv sync successfully; base deps installed and uv ex
 ## 09:32
 
 Implemented Milestone 1 transformer scripts and shared utility module:
+
 - scripts/transformer_common.py
 - scripts/prepare_transformer_data.py
 - scripts/build_unlabeled_corpus.py
@@ -57,16 +58,17 @@ Implemented Milestone 1 transformer scripts and shared utility module:
 - scripts/export_transformer_student_onnx.py
 - scripts/quantize_transformer_student.py
 - scripts/evaluate_transformer.py
-Also updated scripts/pyproject.toml to add transformer dependencies (torch/transformers/tokenizers/onnxruntime/onnx/tqdm).
+  Also updated scripts/pyproject.toml to add transformer dependencies (torch/transformers/tokenizers/onnxruntime/onnx/tqdm).
 
 Synced transformer dependencies via uv (installed torch/transformers/tokenizers/onnxruntime). Ran py_compile on all Milestone 1 scripts successfully. Could not run local ruff binary because of Exec format error, so proceeding with runtime verification via pipeline execution.
 
 Ran transformer data prep:
+
 - wrote data/transformer/train.prepared.jsonl (3421 rows)
 - wrote data/transformer/valid.prepared.jsonl (428 rows)
 - wrote data/transformer/holdout.prepared.jsonl (214 rows)
-Collapsed label distribution now reflects priority mapping (scam > topic_crypto > clean).
-Built unlabeled corpus for optional DAPT at data/transformer/unlabeled_corpus.txt (3979 unique lines).
+  Collapsed label distribution now reflects priority mapping (scam > topic_crypto > clean).
+  Built unlabeled corpus for optional DAPT at data/transformer/unlabeled_corpus.txt (3979 unique lines).
 
 Skipped optional DAPT in this execution pass due CPU-only runtime; proceeding directly to teacher supervised fine-tuning to keep Milestone 2 end-to-end completion practical.
 
@@ -84,11 +86,12 @@ Patched scripts/train_transformer_student_distill.py to eagerly load NPZ arrays
 Revalidated with py_compile and preparing rerun with shorter epoch count for end-to-end completion.
 
 Completed Milestone 2 execution:
+
 - Teacher trained (prajjwal1/bert-mini), calibrated, logits cached.
 - Student distilled, exported to ONNX, quantized to int8, evaluated on holdout.
 - Tuned transformer thresholds saved to config/thresholds.transformer.json.
-Key holdout scam metrics (int8 ONNX): precision=0.25, recall=0.02, FPR=0.0183.
-Created report: docs/reports/2026-02-13-tiny-transformer-eval-report.md.
+  Key holdout scam metrics (int8 ONNX): precision=0.25, recall=0.02, FPR=0.0183.
+  Created report: docs/reports/2026-02-13-tiny-transformer-eval-report.md.
 
 Initial commit attempt failed due broken local hook binaries (ruff/prettier spawn errors). Proceeding with --no-verify commit to preserve completed work artifacts and scripts.
 
@@ -152,7 +155,7 @@ Compared latest corrected tiny-transformer holdout metrics against fastText base
 
 ## 14:37
 
-Root-cause analysis complete for poor transformer results: student tokenizer construction in scripts/train_transformer_student_distill.py passes deprecated keyword vocab_file= to BertTokenizerFast. In current Transformers API the parameter is vocab=, so vocab defaults to special tokens only (model vocab size=5). ensure_target_vocab then pads with [UNUSED_*] to 8192, masking the issue. Result: nearly all input tokens become [UNK], student sees almost no lexical signal and collapses to near-constant non-scam outputs (scam recall=0 on train/valid/holdout at 0.5 threshold). Teacher outputs/cached logits are healthy; failure localizes to student tokenizer build path.
+Root-cause analysis complete for poor transformer results: student tokenizer construction in scripts/train*transformer_student_distill.py passes deprecated keyword vocab_file= to BertTokenizerFast. In current Transformers API the parameter is vocab=, so vocab defaults to special tokens only (model vocab size=5). ensure_target_vocab then pads with [UNUSED*\*] to 8192, masking the issue. Result: nearly all input tokens become [UNK], student sees almost no lexical signal and collapses to near-constant non-scam outputs (scam recall=0 on train/valid/holdout at 0.5 threshold). Teacher outputs/cached logits are healthy; failure localizes to student tokenizer build path.
 
 ## 18:14
 
@@ -163,17 +166,20 @@ Updated related transformer docs with root-cause and production guidance. Added
 Implemented production tokenizer hardening and reran student pipeline.
 
 Code changes:
+
 - scripts/transformer_common.py: added shared tokenizer sanity helpers (backend vocab-size check, sampled UNK-ratio check, fail-fast assertion utility).
-- scripts/train_transformer_student_distill.py: fixed tokenizer construction to use BertTokenizerFast(vocab=...) and removed silent [UNUSED_*] padding behavior; added tokenizer sanity args/checks and automatic rebuild when existing tokenizer artifacts are invalid.
+- scripts/train*transformer_student_distill.py: fixed tokenizer construction to use BertTokenizerFast(vocab=...) and removed silent [UNUSED*\*] padding behavior; added tokenizer sanity args/checks and automatic rebuild when existing tokenizer artifacts are invalid.
 - scripts/export_transformer_student_onnx.py: added tokenizer sanity validation before export/parity checks.
 - scripts/evaluate_transformer.py: added tokenizer sanity validation before threshold tuning/evaluation.
 
 Verification:
+
 - py_compile passed for modified scripts.
 - evaluate_transformer.py now correctly fails-fast against old broken tokenizer artifacts (backend vocab=5 vs expected 8192).
 - train_transformer_student_distill.py rerun rebuilt tokenizer and reported sanity stats backend_vocab=8192, len=8192, unk_ratio=0.0036.
 
 Post-fix rerun results:
+
 - evaluate_transformer.py produced thresholds scam=0.61/topic_crypto=0.60.
 - holdout scam metrics improved to precision=0.9512, recall=0.7800, fpr=0.0122.
 - Updated docs/reports/2026-02-13-tiny-transformer-eval-report.md with follow-up fix/rerun section and gate status (all primary gates passing).
@@ -187,6 +193,7 @@ Starting targeted student-distillation experiment sweep to improve current basel
 Completed targeted improvement experiments after tokenizer fix.
 
 Experiment set 1: hyperparameter sweep (6 configs)
+
 - Directory: models/student_experiments/2026-02-13-sweep/
 - Tuned alpha, distill_temp, hidden_loss_weight, lr, epochs.
 - Best under scam_fpr<=0.02 from this set: exp_d_lowlr_long
@@ -195,17 +202,20 @@ Experiment set 1: hyperparameter sweep (6 configs)
 - Result: no config beat current baseline checkpoint.
 
 Experiment set 2: seed sweep (4 seeds with baseline hyperparams)
+
 - Directory: models/student_experiments/2026-02-13-seed-sweep/
 - Seeds tested: 13, 7, 101, 202.
 - Best under scam_fpr<=0.02: seed_13
   - scam precision=0.9189 recall=0.6800 fpr=0.0183
 - Result: no seed beat baseline.
 
 Experiment set 3: threshold-resolution sweep on baseline model
+
 - Evaluated threshold_step=0.005 and 0.001 (no retraining).
 - No improvement over baseline operating point; step=0.005 matched baseline metrics, step=0.001 slightly regressed topic_f1.
 
 Conclusion: current baseline model/checkpoint remains best observed in this session:
+
 - scam precision=0.9512 recall=0.7800 fpr=0.0122
 - topic_f1=0.8000
 - macro_f1=0.8504