Remove script-centric files from results

results/requirements.txt

@@ -1,11 +0,0 @@
-torch>=2.2.0
-transformers>=4.40.0
-datasets>=2.18.0
-pandas>=2.2.0
-numpy>=1.26.0
-scikit-learn>=1.4.0
-pyarrow>=15.0.0
-fastparquet>=2024.2.0
-joblib>=1.3.0
-xgboost>=2.0.0
-peft>=0.11.0

results/run_end_to_end_bert.md

@@ -1,255 +0,0 @@
-# Run Commands: End-to-End BERT Training
-
-## 1) Go to project folder
-```bash
-cd /Users/jamie/Downloads/LLM_Project
-```
-
-## 2) Activate virtual environment
-```bash
-source .venv/bin/activate
-```
-
-## 3) Train end-to-end BERT (full run)
-```bash
-python bert.py \
-  --model_name boltuix/bert-lite \
-  --data_path enriched_news.parquet \
-  --output_dir outputs_compare_models \
-  --epochs 3 \
-  --train_batch_size 16 \
-  --max_length 128 \
-  --learning_rate 2e-5 \
-  --weight_decay 0.01 \
-  --sample_frac 1.0 \
-  --min_per_stratum 1 \
-  --log_level INFO
-```
-
-## 4) Train end-to-end BERT (faster dev run on laptop)
-```bash
-python bert.py \
-  --model_name boltuix/bert-lite \
-  --data_path enriched_news.parquet \
-  --output_dir outputs_compare_models \
-  --fast_mode \
-  --sample_frac 0.35 \
-  --min_per_stratum 50 \
-  --log_level INFO
-```
-
-## 5) Train with LoRA (faster and lighter than full fine-tuning)
-```bash
-python bert.py \
-  --model_name kk08/CryptoBERT \
-  --data_path enriched_news.parquet \
-  --output_dir outputs_compare_models \
-  --peft_mode lora \
-  --lora_r 8 \
-  --lora_alpha 16 \
-  --lora_dropout 0.05 \
-  --lora_target_modules query,key,value \
-  --train_batch_size 8 \
-  --grad_accum_steps 2 \
-  --num_workers 2 \
-  --eval_every_epochs 1 \
-  --sample_frac 0.5 \
-  --min_per_stratum 50 \
-  --log_level INFO
-```
-
-## 6) Extra speed-focused run (LoRA + fast mode)
-```bash
-python bert.py \
-  --model_name boltuix/bert-lite \
-  --data_path enriched_news.parquet \
-  --output_dir outputs_compare_models \
-  --peft_mode lora \
-  --fast_mode \
-  --grad_accum_steps 2 \
-  --num_workers 2 \
-  --sample_frac 0.35 \
-  --min_per_stratum 50 \
-  --log_level INFO
-```
-
-## 7) Output files to check
-- `outputs_compare_models/best_neural_model.pt`
-- `outputs_compare_models/metrics_summary_neural.json`
-- `outputs_compare_models/numeric_scaler.joblib`
-- `outputs_compare_models/fng_onehot_encoder.joblib`
-
-## 8) Optional: freeze BERT (not end-to-end fine-tuning)
-```bash
-python bert.py --freeze_bert
-```
-
-## 9) Train XGBoost only (frozen CLS + numeric)
-```bash
-python xgb.py \
-  --model_name gaunernst/bert-mini-uncased \
-  --data_path enriched_news.parquet \
-  --output_json outputs_compare_models/metrics_xgb.json \
-  --sample_frac 1.0 \
-  --min_per_stratum 1 \
-  --xgb_random_search \
-  --xgb_random_iters 20 \
-  --xgb_refit_on_train_val \
-  --log_level INFO
-```
-
-## 10) Train XGBoost only (faster run)
-```bash
-python xgb.py \
-  --model_name gaunernst/bert-mini-uncased \
-  --data_path enriched_news.parquet \
-  --sample_frac 0.35 \
-  --min_per_stratum 50 \
-  --embedding_batch_size 32 \
-  --max_length 96 \
-  --xgb_n_estimators 200 \
-  --xgb_max_depth 4 \
-  --xgb_random_search \
-  --xgb_random_iters 10 \
-  --xgb_refit_on_train_val \
-  --log_level INFO
-```
-
-## 11) Compare numeric-only XGB vs frozen FinBERT CLS+numeric XGB
-```bash
-python compare_xgb_text_vs_numeric.py \
-  --embedding_source frozen \
-  --model_name ProsusAI/finbert \
-  --data_path enriched_news.parquet \
-  --sample_frac 1.0 \
-  --min_per_stratum 1 \
-  --xgb_random_search \
-  --xgb_random_iters 20 \
-  --xgb_refit_on_train_val \
-  --log_level INFO
-```
-
-## 12) Compare numeric-only XGB vs fine-tuned CLS+numeric XGB
-```bash
-python compare_xgb_text_vs_numeric.py \
-  --embedding_source finetuned \
-  --model_name boltuix/bert-lite \
-  --finetuned_checkpoint outputs_compare_models/best_neural_model_base_boltuix_bert-lite.pt \
-  --tokenizer_path outputs_compare_models \
-  --data_path enriched_news.parquet \
-  --sample_frac 1.0 \
-  --min_per_stratum 1 \
-  --xgb_random_search \
-  --xgb_random_iters 20 \
-  --xgb_refit_on_train_val \
-  --log_level INFO
-```
-
-## 13) Compare script (faster run)
-```bash
-python compare_xgb_text_vs_numeric.py \
-  --embedding_source frozen \
-  --model_name ProsusAI/finbert \
-  --sample_frac 0.35 \
-  --min_per_stratum 50 \
-  --embedding_batch_size 32 \
-  --max_length 96 \
-  --xgb_random_search \
-  --xgb_random_iters 10 \
-  --xgb_refit_on_train_val \
-  --log_level INFO
-```
-
-## 14) Typical workflow (recommended order)
-```bash
-# Step A: train neural model (full or LoRA)
-python bert.py --peft_mode lora --fast_mode --sample_frac 0.35 --min_per_stratum 50 --log_level INFO
-
-# Step B: compare numeric-only vs fine-tuned CLS+numeric
-python compare_xgb_text_vs_numeric.py --finetuned_checkpoint outputs_compare_models/best_neural_model.pt --tokenizer_path outputs_compare_models --sample_frac 0.35 --min_per_stratum 50 --log_level INFO
-
-# Optional Step C: run frozen-CLS XGB baseline script directly
-python xgb.py --sample_frac 0.35 --min_per_stratum 50 --xgb_random_search --xgb_refit_on_train_val --log_level INFO
-```
-
-## 15) Useful output files from other scripts
-- `outputs_compare_models/metrics_xgb.json` (from `xgb.py`)
-- `outputs_compare_models/metrics_xgb_cls_vs_numeric.json` (from `compare_xgb_text_vs_numeric.py`)
-
-## 16) Embedding cache safety (important)
-- Caches now use an auto key derived from model + sample settings + seed + max length.
-- This makes cache reuse safer when you change `sample_frac`, model, or other settings.
-- Optional manual override (compare script):
-```bash
-python compare_xgb_text_vs_numeric.py \
-  --embedding_source frozen \
-  --model_name ProsusAI/finbert \
-  --embedding_cache_key my_custom_cache_v1
-```
-- If you want strict isolation per experiment, use a separate cache folder too:
-```bash
-python compare_xgb_text_vs_numeric.py \
-  --embedding_source frozen \
-  --model_name ProsusAI/finbert \
-  --embedding_cache_dir embedding_cache_xgb_compare_exp1
-```
-
-# After getting correct embeddings (this takes a hour to run!!!)
-```bash
-python compare_xgb_text_vs_numeric.py \
-  --embedding_source frozen \
-  --model_name ProsusAI/finbert \
-  --data_path enriched_news.parquet \
-  --sample_frac 1.0 \
-  --min_per_stratum 1 \
-  --embedding_cache_dir embedding_cache_xgb_compare_fresh_20260316 \
-  --embedding_cache_key finbert_full_fresh \
-  --xgb_random_iters 1 \
-  --xgb_n_estimators 100 \
-  --xgb_max_depth 3 \
-  --log_level INFO
-```
-
-Copy-safe single line:
-```bash
-python compare_xgb_text_vs_numeric.py --embedding_source frozen --model_name ProsusAI/finbert --data_path enriched_news.parquet --sample_frac 1.0 --min_per_stratum 1 --embedding_cache_dir embedding_cache_xgb_compare_fresh_20260316 --embedding_cache_key finbert_full_fresh --xgb_random_iters 1 --xgb_n_estimators 100 --xgb_max_depth 3 --log_level INFO
-```
-
-# Fast version of frozen cls (bert-lite) - similar performance against FinBert
-```bash
-python compare_xgb_text_vs_numeric.py \
-  --embedding_source frozen \
-  --model_name boltuix/bert-lite \
-  --data_path enriched_news.parquet \
-  --sample_frac 1.0 \
-  --min_per_stratum 1 \
-  --embedding_cache_dir embedding_cache_xgb_compare_fresh_20260316 \
-  --embedding_cache_key bertlite_full_fresh \
-  --log_level INFO 
-```
-
-## 17) Reproducible run settings (recommended)
-- Use fixed `--seed` and fixed cache key.
-- Use `--deterministic_run` (forces single-thread XGBoost/CV workers).
-- Keep data/sampling settings unchanged across reruns.
-
-```bash
-python compare_xgb_text_vs_numeric.py \
-  --embedding_source frozen \
-  --model_name boltuix/bert-lite \
-  --data_path enriched_news.parquet \
-  --sample_frac 1.0 \
-  --min_per_stratum 1 \
-  --seed 42 \
-  --embedding_cache_dir embedding_cache_xgb_compare_fresh_20260316 \
-  --embedding_cache_key bertlite_full_fresh \
-  --deterministic_run \
-  --xgb_random_iters 20 \
-  --xgb_refit_on_train_val \
-  --log_level INFO
-```
-
-Optional: if you do not use `--deterministic_run`, set workers manually:
-```bash
-python compare_xgb_text_vs_numeric.py --xgb_n_jobs 1
-```