Upload folder using huggingface_hub

10_paper_strengthening/README.md

@@ -0,0 +1,99 @@
+# Paper Strengthening Experiments
+
+These scripts add the experiments needed to make the project closer to a
+defensible paper rather than only a working demo.
+
+Run them after `01_data`, `02_predictor`, and the first `04_diffusion`
+conditioned model are complete.
+
+Before running any script on the GPU server, set the project-local paths:
+
+```bash
+cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/genrl-enhancer-diffusion
+
+export PROJECT_ROOT=$(pwd)
+export RUN_ROOT=${PROJECT_ROOT}/paper_runs
+export HF_ENDPOINT=https://hf-mirror.com
+export TRANSFORMERS_NO_TF=1
+export USE_TF=0
+export TOKENIZERS_PARALLELISM=false
+
+source scripts/00_setup/env.sh
+
+export DEEPSTARR_DIR=${PROJECT_ROOT}/datas/DeepSTARR-enhancer-activity
+export DEEPSTARR_DATASET_ID=${DEEPSTARR_DIR}
+export GENERATOR_BASE_MODEL=${PROJECT_ROOT}/models/GENERator-eukaryote-1.2b-base
+export GENERANNO_BASE_MODEL=${PROJECT_ROOT}/models/GENERanno-eukaryote-0.5b-base
+export PREDICTOR_DIR=${PROJECT_ROOT}/paper_runs/results/deepstarr_regression
+export PREDICTOR_MODEL=${PREDICTOR_DIR}/best_model
+export DIFFUSION_MODEL=${PROJECT_ROOT}/saved_model/deepstarr_discrete_diffusion
+```
+
+## Priority Order
+
+### P0: Minimum paper closure
+
+Run AR baselines, score them with the same predictor, score diffusion with the
+same predictor, then build tables and figures.
+
+```bash
+nohup bash scripts/10_paper_strengthening/p0_minimum_paper_closure.sh > p0_minimum_paper_closure.log 2>&1 &
+tail -f p0_minimum_paper_closure.log
+```
+
+This is the minimum set for a paper-style comparison:
+
+- Reference DeepSTARR sequences.
+- GC-matched random baseline.
+- AR unconditional generation.
+- AR bucket-conditioned generation.
+- Masked diffusion bucket-conditioned generation.
+- Shared predictor scoring.
+- Sequence quality and distribution metrics.
+
+### P1: Core ablations
+
+Mask-ratio ablation:
+
+```bash
+nohup bash scripts/10_paper_strengthening/p1_mask_ratio_ablation.sh > p1_mask_ratio_ablation.log 2>&1 &
+tail -f p1_mask_ratio_ablation.log
+```
+
+Conditioning ablation:
+
+```bash
+nohup bash scripts/10_paper_strengthening/p1_conditioning_ablation.sh > p1_conditioning_ablation.log 2>&1 &
+tail -f p1_conditioning_ablation.log
+```
+
+These directly support the paper claim that mask-controlled diffusion is a
+mutation-budget design method and that conditioning improves controllability.
+
+### P2: Final split and efficiency analysis
+
+Evaluate the best diffusion model on the test split:
+
+```bash
+nohup bash scripts/10_paper_strengthening/p2_test_split_diffusion.sh > p2_test_split_diffusion.log 2>&1 &
+tail -f p2_test_split_diffusion.log
+```
+
+Evaluate different denoising step counts:
+
+```bash
+nohup bash scripts/10_paper_strengthening/p2_diffusion_steps_ablation.sh > p2_diffusion_steps_ablation.log 2>&1 &
+tail -f p2_diffusion_steps_ablation.log
+```
+
+### P3: Seed robustness
+
+Run generation/evaluation with several seeds:
+
+```bash
+nohup bash scripts/10_paper_strengthening/p3_seed_repeats_diffusion.sh > p3_seed_repeats_diffusion.log 2>&1 &
+tail -f p3_seed_repeats_diffusion.log
+```
+
+Use the resulting means and standard deviations in the final tables.
+

10_paper_strengthening/p0_minimum_paper_closure.sh

@@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+: "${DEEPSTARR_DIR:?Set DEEPSTARR_DIR to the local DeepSTARR parquet directory.}"
+: "${GENERATOR_BASE_MODEL:?Set GENERATOR_BASE_MODEL to the local GENERator model.}"
+: "${GENERANNO_BASE_MODEL:?Set GENERANNO_BASE_MODEL to the local GENERanno model.}"
+: "${PREDICTOR_MODEL:?Set PREDICTOR_MODEL to the trained predictor best_model directory.}"
+: "${DIFFUSION_MODEL:?Set DIFFUSION_MODEL to the trained diffusion saved model directory.}"
+
+export NUM_PER_BUCKET="${NUM_PER_BUCKET:-128}"
+export NUM_DIFFUSION_STEPS="${NUM_DIFFUSION_STEPS:-64}"
+export DIFFUSION_EVAL_BATCH_SIZE="${DIFFUSION_EVAL_BATCH_SIZE:-512}"
+export PREDICTOR_SCORE_BATCH_SIZE="${PREDICTOR_SCORE_BATCH_SIZE:-64}"
+export DIFFUSION_SCORE_BATCH_SIZE="${DIFFUSION_SCORE_BATCH_SIZE:-64}"
+export PLL_CHUNK_SIZE="${PLL_CHUNK_SIZE:-64}"
+
+bash "${SCRIPT_DIR}/../02_predictor/plot_predictor_validation.sh"
+
+bash "${SCRIPT_DIR}/../03_ar_generation/train_ar_unconditional.sh"
+bash "${SCRIPT_DIR}/../03_ar_generation/train_ar_conditioned.sh"
+bash "${SCRIPT_DIR}/../03_ar_generation/evaluate_ar_unconditional.sh"
+bash "${SCRIPT_DIR}/../03_ar_generation/evaluate_ar_conditioned.sh"
+
+bash "${SCRIPT_DIR}/../05_scoring/score_ar_unconditional.sh"
+bash "${SCRIPT_DIR}/../05_scoring/score_ar_conditioned.sh"
+
+bash "${SCRIPT_DIR}/../04_diffusion/evaluate_diffusion_with_predictor.sh"
+
+bash "${SCRIPT_DIR}/../06_sequence_metrics/compute_sequence_metrics.sh"
+bash "${SCRIPT_DIR}/../08_visualization/make_paper_figures.sh"
+
+echo "P0 minimum paper closure completed under ${RUN_ROOT}"
+

10_paper_strengthening/p1_conditioning_ablation.sh

@@ -0,0 +1,65 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+: "${DEEPSTARR_DIR:?Set DEEPSTARR_DIR to the local DeepSTARR parquet directory.}"
+: "${GENERANNO_BASE_MODEL:?Set GENERANNO_BASE_MODEL to the local GENERanno model.}"
+: "${PREDICTOR_MODEL:?Set PREDICTOR_MODEL to the trained predictor best_model directory.}"
+
+export DIFFUSION_ABLATION_EPOCHS="${DIFFUSION_ABLATION_EPOCHS:-1}"
+export DIFFUSION_ABLATION_BATCH_SIZE="${DIFFUSION_ABLATION_BATCH_SIZE:-128}"
+export DIFFUSION_ABLATION_GRAD_ACCUM="${DIFFUSION_ABLATION_GRAD_ACCUM:-1}"
+export DIFFUSION_ABLATION_LR="${DIFFUSION_ABLATION_LR:-5e-5}"
+export DIFFUSION_EVAL_BATCH_SIZE="${DIFFUSION_EVAL_BATCH_SIZE:-512}"
+export PREDICTOR_SCORE_BATCH_SIZE="${PREDICTOR_SCORE_BATCH_SIZE:-64}"
+export NUM_PER_BUCKET="${NUM_PER_BUCKET:-128}"
+export NUM_DIFFUSION_STEPS="${NUM_DIFFUSION_STEPS:-64}"
+
+mkdir -p "${MODEL_ROOT}/ablations" "${RESULT_ROOT}/ablations"
+
+cd "${GENERANNO_DIR}"
+
+UNCOND_MODEL="${MODEL_ROOT}/ablations/diffusion_unconditioned"
+UNCOND_CKPT="${RESULT_ROOT}/checkpoints/diffusion_unconditioned"
+UNCOND_RESULT="${RESULT_ROOT}/ablations/diffusion_unconditioned_valid_predictor"
+
+python3 src/tasks/downstream/discrete_diffusion_train.py \
+  --model_name "${GENERANNO_BASE_MODEL}" \
+  --dataset_dir "${DEEPSTARR_DIR}" \
+  --output_dir "${UNCOND_CKPT}" \
+  --saved_model_dir "${UNCOND_MODEL}" \
+  --sequence_col sequence \
+  --label_col label \
+  --score_mode sum \
+  --max_length "${DIFFUSION_MAX_LENGTH:-256}" \
+  --num_train_epochs "${DIFFUSION_ABLATION_EPOCHS}" \
+  --batch_size "${DIFFUSION_ABLATION_BATCH_SIZE}" \
+  --gradient_accumulation_steps "${DIFFUSION_ABLATION_GRAD_ACCUM}" \
+  --learning_rate "${DIFFUSION_ABLATION_LR}" \
+  --num_diffusion_steps "${NUM_DIFFUSION_STEPS}" \
+  --bf16 \
+  --gradient_checkpointing \
+  --report_to none \
+  --run_name diffusion_unconditioned
+
+python3 src/tasks/downstream/discrete_diffusion_evaluate.py \
+  --diffusion_model "${UNCOND_MODEL}" \
+  --base_model_for_code "${GENERANNO_BASE_MODEL}" \
+  --dataset_dir "${DEEPSTARR_DIR}" \
+  --predictor_model "${PREDICTOR_MODEL}" \
+  --split valid \
+  --num_per_bucket "${NUM_PER_BUCKET}" \
+  --sequence_length "${SEQUENCE_LENGTH:-246}" \
+  --num_diffusion_steps "${NUM_DIFFUSION_STEPS}" \
+  --batch_size "${DIFFUSION_EVAL_BATCH_SIZE}" \
+  --predictor_batch_size "${PREDICTOR_SCORE_BATCH_SIZE}" \
+  --max_length "${DIFFUSION_MAX_LENGTH:-256}" \
+  --temperature "${DIFFUSION_TEMPERATURE:-1.0}" \
+  --bf16 \
+  --attn_implementation "${ATTN_IMPLEMENTATION:-sdpa}" \
+  --output_dir "${UNCOND_RESULT}"
+
+echo "P1 conditioning ablation completed under ${UNCOND_RESULT}"
+

10_paper_strengthening/p1_mask_ratio_ablation.sh

@@ -0,0 +1,79 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+: "${DEEPSTARR_DIR:?Set DEEPSTARR_DIR to the local DeepSTARR parquet directory.}"
+: "${GENERANNO_BASE_MODEL:?Set GENERANNO_BASE_MODEL to the local GENERanno model.}"
+: "${PREDICTOR_MODEL:?Set PREDICTOR_MODEL to the trained predictor best_model directory.}"
+
+export DIFFUSION_ABLATION_EPOCHS="${DIFFUSION_ABLATION_EPOCHS:-1}"
+export DIFFUSION_ABLATION_BATCH_SIZE="${DIFFUSION_ABLATION_BATCH_SIZE:-1024}"
+export DIFFUSION_ABLATION_GRAD_ACCUM="${DIFFUSION_ABLATION_GRAD_ACCUM:-1}"
+export DIFFUSION_ABLATION_LR="${DIFFUSION_ABLATION_LR:-5e-5}"
+export DIFFUSION_EVAL_BATCH_SIZE="${DIFFUSION_EVAL_BATCH_SIZE:-1024}"
+export PREDICTOR_SCORE_BATCH_SIZE="${PREDICTOR_SCORE_BATCH_SIZE:-1024}"
+export NUM_PER_BUCKET="${NUM_PER_BUCKET:-128}"
+export NUM_DIFFUSION_STEPS="${NUM_DIFFUSION_STEPS:-64}"
+
+mkdir -p "${MODEL_ROOT}/ablations" "${RESULT_ROOT}/ablations"
+
+cd "${GENERANNO_DIR}"
+
+run_mask_ablation() {
+  local name="$1"
+  local mask_min="$2"
+  local mask_max="$3"
+  local ckpt_dir="${RESULT_ROOT}/checkpoints/diffusion_mask_${name}"
+  local model_dir="${MODEL_ROOT}/ablations/diffusion_mask_${name}"
+  local result_dir="${RESULT_ROOT}/ablations/diffusion_mask_${name}_valid_predictor"
+
+  python3 src/tasks/downstream/discrete_diffusion_train.py \
+    --model_name "${GENERANNO_BASE_MODEL}" \
+    --dataset_dir "${DEEPSTARR_DIR}" \
+    --output_dir "${ckpt_dir}" \
+    --saved_model_dir "${model_dir}" \
+    --conditioned \
+    --sequence_col sequence \
+    --label_col label \
+    --score_mode sum \
+    --max_length "${DIFFUSION_MAX_LENGTH:-256}" \
+    --num_train_epochs "${DIFFUSION_ABLATION_EPOCHS}" \
+    --batch_size "${DIFFUSION_ABLATION_BATCH_SIZE}" \
+    --gradient_accumulation_steps "${DIFFUSION_ABLATION_GRAD_ACCUM}" \
+    --learning_rate "${DIFFUSION_ABLATION_LR}" \
+    --mask_prob_min "${mask_min}" \
+    --mask_prob_max "${mask_max}" \
+    --num_diffusion_steps "${NUM_DIFFUSION_STEPS}" \
+    --bf16 \
+    --gradient_checkpointing \
+    --report_to none \
+    --run_name "diffusion_mask_${name}"
+
+  python3 src/tasks/downstream/discrete_diffusion_evaluate.py \
+    --diffusion_model "${model_dir}" \
+    --base_model_for_code "${GENERANNO_BASE_MODEL}" \
+    --dataset_dir "${DEEPSTARR_DIR}" \
+    --predictor_model "${PREDICTOR_MODEL}" \
+    --split valid \
+    --conditioned \
+    --num_per_bucket "${NUM_PER_BUCKET}" \
+    --sequence_length "${SEQUENCE_LENGTH:-246}" \
+    --num_diffusion_steps "${NUM_DIFFUSION_STEPS}" \
+    --batch_size "${DIFFUSION_EVAL_BATCH_SIZE}" \
+    --predictor_batch_size "${PREDICTOR_SCORE_BATCH_SIZE}" \
+    --max_length "${DIFFUSION_MAX_LENGTH:-256}" \
+    --temperature "${DIFFUSION_TEMPERATURE:-1.0}" \
+    --bf16 \
+    --attn_implementation "${ATTN_IMPLEMENTATION:-sdpa}" \
+    --output_dir "${result_dir}"
+}
+
+run_mask_ablation "015_030" 0.15 0.30
+run_mask_ablation "030_050" 0.30 0.50
+run_mask_ablation "050_070" 0.50 0.70
+run_mask_ablation "070_090" 0.70 0.90
+
+echo "P1 mask-ratio ablation completed under ${RESULT_ROOT}/ablations"
+

10_paper_strengthening/p2_diffusion_steps_ablation.sh

@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+: "${DEEPSTARR_DIR:?Set DEEPSTARR_DIR to the local DeepSTARR parquet directory.}"
+: "${DIFFUSION_MODEL:?Set DIFFUSION_MODEL to the trained diffusion saved model directory.}"
+: "${GENERANNO_BASE_MODEL:?Set GENERANNO_BASE_MODEL to the local GENERanno model.}"
+: "${PREDICTOR_MODEL:?Set PREDICTOR_MODEL to the trained predictor best_model directory.}"
+
+export NUM_PER_BUCKET="${NUM_PER_BUCKET:-128}"
+export DIFFUSION_EVAL_BATCH_SIZE="${DIFFUSION_EVAL_BATCH_SIZE:-512}"
+export PREDICTOR_SCORE_BATCH_SIZE="${PREDICTOR_SCORE_BATCH_SIZE:-64}"
+
+mkdir -p "${RESULT_ROOT}/ablations"
+
+cd "${GENERANNO_DIR}"
+
+for steps in 16 32 64 128; do
+  python3 src/tasks/downstream/discrete_diffusion_evaluate.py \
+    --diffusion_model "${DIFFUSION_MODEL}" \
+    --base_model_for_code "${GENERANNO_BASE_MODEL}" \
+    --dataset_dir "${DEEPSTARR_DIR}" \
+    --predictor_model "${PREDICTOR_MODEL}" \
+    --split valid \
+    --conditioned \
+    --num_per_bucket "${NUM_PER_BUCKET}" \
+    --sequence_length "${SEQUENCE_LENGTH:-246}" \
+    --num_diffusion_steps "${steps}" \
+    --batch_size "${DIFFUSION_EVAL_BATCH_SIZE}" \
+    --predictor_batch_size "${PREDICTOR_SCORE_BATCH_SIZE}" \
+    --max_length "${DIFFUSION_MAX_LENGTH:-256}" \
+    --temperature "${DIFFUSION_TEMPERATURE:-1.0}" \
+    --bf16 \
+    --attn_implementation "${ATTN_IMPLEMENTATION:-sdpa}" \
+    --output_dir "${RESULT_ROOT}/ablations/diffusion_steps_${steps}_valid_predictor"
+done
+
+echo "P2 diffusion step ablation completed under ${RESULT_ROOT}/ablations"
+

10_paper_strengthening/p2_test_split_diffusion.sh

@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+: "${DEEPSTARR_DIR:?Set DEEPSTARR_DIR to the local DeepSTARR parquet directory.}"
+: "${DIFFUSION_MODEL:?Set DIFFUSION_MODEL to the trained diffusion saved model directory.}"
+: "${GENERANNO_BASE_MODEL:?Set GENERANNO_BASE_MODEL to the local GENERanno model.}"
+: "${PREDICTOR_MODEL:?Set PREDICTOR_MODEL to the trained predictor best_model directory.}"
+
+export NUM_PER_BUCKET="${NUM_PER_BUCKET:-128}"
+export NUM_DIFFUSION_STEPS="${NUM_DIFFUSION_STEPS:-64}"
+export DIFFUSION_EVAL_BATCH_SIZE="${DIFFUSION_EVAL_BATCH_SIZE:-512}"
+export PREDICTOR_SCORE_BATCH_SIZE="${PREDICTOR_SCORE_BATCH_SIZE:-64}"
+
+cd "${GENERANNO_DIR}"
+python3 src/tasks/downstream/discrete_diffusion_evaluate.py \
+  --diffusion_model "${DIFFUSION_MODEL}" \
+  --base_model_for_code "${GENERANNO_BASE_MODEL}" \
+  --dataset_dir "${DEEPSTARR_DIR}" \
+  --predictor_model "${PREDICTOR_MODEL}" \
+  --split test \
+  --conditioned \
+  --num_per_bucket "${NUM_PER_BUCKET}" \
+  --sequence_length "${SEQUENCE_LENGTH:-246}" \
+  --num_diffusion_steps "${NUM_DIFFUSION_STEPS}" \
+  --batch_size "${DIFFUSION_EVAL_BATCH_SIZE}" \
+  --predictor_batch_size "${PREDICTOR_SCORE_BATCH_SIZE}" \
+  --max_length "${DIFFUSION_MAX_LENGTH:-256}" \
+  --temperature "${DIFFUSION_TEMPERATURE:-1.0}" \
+  --bf16 \
+  --attn_implementation "${ATTN_IMPLEMENTATION:-sdpa}" \
+  --output_dir "${RESULT_ROOT}/diffusion_test_predictor"
+
+echo "P2 test split diffusion evaluation completed under ${RESULT_ROOT}/diffusion_test_predictor"
+

10_paper_strengthening/p3_seed_repeats_diffusion.sh

@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+: "${DEEPSTARR_DIR:?Set DEEPSTARR_DIR to the local DeepSTARR parquet directory.}"
+: "${DIFFUSION_MODEL:?Set DIFFUSION_MODEL to the trained diffusion saved model directory.}"
+: "${GENERANNO_BASE_MODEL:?Set GENERANNO_BASE_MODEL to the local GENERanno model.}"
+: "${PREDICTOR_MODEL:?Set PREDICTOR_MODEL to the trained predictor best_model directory.}"
+
+export NUM_PER_BUCKET="${NUM_PER_BUCKET:-128}"
+export NUM_DIFFUSION_STEPS="${NUM_DIFFUSION_STEPS:-64}"
+export DIFFUSION_EVAL_BATCH_SIZE="${DIFFUSION_EVAL_BATCH_SIZE:-512}"
+export PREDICTOR_SCORE_BATCH_SIZE="${PREDICTOR_SCORE_BATCH_SIZE:-64}"
+export DIFFUSION_SEEDS="${DIFFUSION_SEEDS:-11 22 33}"
+
+mkdir -p "${RESULT_ROOT}/seed_repeats"
+
+cd "${GENERANNO_DIR}"
+
+for seed in ${DIFFUSION_SEEDS}; do
+  python3 src/tasks/downstream/discrete_diffusion_evaluate.py \
+    --diffusion_model "${DIFFUSION_MODEL}" \
+    --base_model_for_code "${GENERANNO_BASE_MODEL}" \
+    --dataset_dir "${DEEPSTARR_DIR}" \
+    --predictor_model "${PREDICTOR_MODEL}" \
+    --split valid \
+    --conditioned \
+    --num_per_bucket "${NUM_PER_BUCKET}" \
+    --sequence_length "${SEQUENCE_LENGTH:-246}" \
+    --num_diffusion_steps "${NUM_DIFFUSION_STEPS}" \
+    --batch_size "${DIFFUSION_EVAL_BATCH_SIZE}" \
+    --predictor_batch_size "${PREDICTOR_SCORE_BATCH_SIZE}" \
+    --max_length "${DIFFUSION_MAX_LENGTH:-256}" \
+    --temperature "${DIFFUSION_TEMPERATURE:-1.0}" \
+    --seed "${seed}" \
+    --bf16 \
+    --attn_implementation "${ATTN_IMPLEMENTATION:-sdpa}" \
+    --output_dir "${RESULT_ROOT}/seed_repeats/diffusion_seed_${seed}_valid_predictor"
+done
+
+echo "P3 diffusion seed repeats completed under ${RESULT_ROOT}/seed_repeats"
+

evaluate_ar_conditioned.sh

@@ -0,0 +1,27 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+for filename in modeling_generator.py tokenizer.py; do
+  if [[ -f "${GENERATOR_BASE_MODEL}/${filename}" && ! -f "${AR_COND_MODEL}/${filename}" ]]; then
+    cp "${GENERATOR_BASE_MODEL}/${filename}" "${AR_COND_MODEL}/${filename}"
+  fi
+done
+
+cd "${GENERATOR_DIR}"
+python3 src/tasks/downstream/generation_validation.py \
+  --model_name "${AR_COND_MODEL}" \
+  --parquet_path "${CONDITIONED_DEEPSTARR_DIR}/valid.parquet" \
+  --sequence_col conditioned_sequence \
+  --conditioned_input \
+  --output_dir "${RESULT_ROOT}/ar_conditioned_valid" \
+  --num_samples "${NUM_GENERATION_SAMPLES:-384}" \
+  --prompt_bp_length "${PROMPT_BP_LENGTH:-120}" \
+  --continuation_bp_length "${CONTINUATION_BP_LENGTH:-126}" \
+  --batch_size "${GEN_BATCH_SIZE:-512}" \
+  --attn_implementation "${ATTN_IMPLEMENTATION:-sdpa}" \
+  --bf16 \
+  --report_to none \
+  --run_name ar_conditioned_valid

evaluate_ar_unconditional.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+for filename in modeling_generator.py tokenizer.py; do
+  if [[ -f "${GENERATOR_BASE_MODEL}/${filename}" && ! -f "${AR_UNCOND_MODEL}/${filename}" ]]; then
+    cp "${GENERATOR_BASE_MODEL}/${filename}" "${AR_UNCOND_MODEL}/${filename}"
+  fi
+done
+
+cd "${GENERATOR_DIR}"
+python3 src/tasks/downstream/generation_validation.py \
+  --model_name "${AR_UNCOND_MODEL}" \
+  --parquet_path "${DEEPSTARR_DIR}/valid.parquet" \
+  --sequence_col sequence \
+  --output_dir "${RESULT_ROOT}/ar_unconditional_valid" \
+  --num_samples "${NUM_GENERATION_SAMPLES:-384}" \
+  --prompt_bp_length "${PROMPT_BP_LENGTH:-120}" \
+  --continuation_bp_length "${CONTINUATION_BP_LENGTH:-126}" \
+  --batch_size "${GEN_BATCH_SIZE:-512}" \
+  --attn_implementation "${ATTN_IMPLEMENTATION:-sdpa}" \
+  --bf16 \
+  --report_to none \
+  --run_name ar_unconditional_valid

train_ar_conditioned.sh

@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+cd "${GENERATOR_DIR}"
+python3 src/tasks/downstream/fine_tuning.py \
+  --model_name "${GENERATOR_BASE_MODEL}" \
+  --parquet_path "${CONDITIONED_DEEPSTARR_DIR}/train.parquet" \
+  --sequence_col conditioned_sequence \
+  --conditioned_input \
+  --output_dir "${RESULT_ROOT}/checkpoints/deepstarr_sft_conditioned" \
+  --saved_model_dir "${AR_COND_MODEL}" \
+  --tmp_dir "${RUN_ROOT}/tmp/deepstarr_sft_conditioned" \
+  --epochs "${AR_EPOCHS:-3}" \
+  --batch_size "${AR_BATCH_SIZE:-4}" \
+  --gradient_accumulation "${AR_GRAD_ACCUM:-1}" \
+  --lr "${AR_LR:-5e-5}" \
+  --max_token_length "${AR_MAX_TOKEN_LENGTH:-256}" \
+  --attn_implementation "${ATTN_IMPLEMENTATION:-sdpa}" \
+  --bf16 \
+  --report_to none \
+  --run_name deepstarr_sft_conditioned
+
+for filename in modeling_generator.py tokenizer.py; do
+  if [[ -f "${GENERATOR_BASE_MODEL}/${filename}" && ! -f "${AR_COND_MODEL}/${filename}" ]]; then
+    cp "${GENERATOR_BASE_MODEL}/${filename}" "${AR_COND_MODEL}/${filename}"
+  fi
+done

train_ar_unconditional.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "${SCRIPT_DIR}/../00_setup/env.sh"
+
+cd "${GENERATOR_DIR}"
+python3 src/tasks/downstream/fine_tuning.py \
+  --model_name "${GENERATOR_BASE_MODEL}" \
+  --parquet_path "${DEEPSTARR_DIR}/train.parquet" \
+  --sequence_col sequence \
+  --output_dir "${RESULT_ROOT}/checkpoints/deepstarr_sft" \
+  --saved_model_dir "${AR_UNCOND_MODEL}" \
+  --tmp_dir "${RUN_ROOT}/tmp/deepstarr_sft" \
+  --epochs "${AR_EPOCHS:-3}" \
+  --batch_size "${AR_BATCH_SIZE:-4}" \
+  --gradient_accumulation "${AR_GRAD_ACCUM:-1}" \
+  --lr "${AR_LR:-5e-5}" \
+  --max_token_length "${AR_MAX_TOKEN_LENGTH:-256}" \
+  --attn_implementation "${ATTN_IMPLEMENTATION:-sdpa}" \
+  --bf16 \
+  --report_to none \
+  --run_name deepstarr_sft_unconditional
+
+for filename in modeling_generator.py tokenizer.py; do
+  if [[ -f "${GENERATOR_BASE_MODEL}/${filename}" && ! -f "${AR_UNCOND_MODEL}/${filename}" ]]; then
+    cp "${GENERATOR_BASE_MODEL}/${filename}" "${AR_UNCOND_MODEL}/${filename}"
+  fi
+done