From 29719f4d19e0411d767299b1cc763c34fc41aa55 Mon Sep 17 00:00:00 2001
From: 51616 <rujikorn.ch@gmail.com>
Date: Tue, 29 Jul 2025 06:59:58 +0000
Subject: [PATCH] repeat_prompts + faster eval

---
 src/ctx_to_lora/configs.py          |  4 +++
 src/ctx_to_lora/data/definitions.py | 47 ++++++++++++++---------------
 src/ctx_to_lora/eval_utils.py       |  9 ++++--
 3 files changed, 33 insertions(+), 27 deletions(-)

diff --git a/src/ctx_to_lora/configs.py b/src/ctx_to_lora/configs.py
index 47c3b56..ac28238 100644
--- a/src/ctx_to_lora/configs.py
+++ b/src/ctx_to_lora/configs.py
@@ -144,6 +144,10 @@ class TrainingArguments(TrainingArguments):
             "Options: 'inputs', 'loss'."
         },
     )
+    average_tokens_across_devices: bool = field(
+        default=False,
+        metadata={"help": "compute num_items_in_batch across devices."},
+    )
     # mem leak if use persistent workers
     # https://github.com/pytorch/pytorch/issues/62066
     # https://github.com/huggingface/transformers/issues/30943
diff --git a/src/ctx_to_lora/data/definitions.py b/src/ctx_to_lora/data/definitions.py
index 489e42f..b9e653e 100644
--- a/src/ctx_to_lora/data/definitions.py
+++ b/src/ctx_to_lora/data/definitions.py
@@ -22,30 +22,29 @@ SELF_GEN_SYSTEM_MSG = (
 
 
 REPEAT_PROMPTS = [
-    "Repeat the text.",
-    "Repeat the text above.",
-    "Repeat the above text.",
-    "Repeat exactly what was written.",
-    "Duplicate the text.",
-    "Echo the text. Do not output other words.",
-    "Repeat the content provided.",
-    "Copy the content. Reply with just the content.",
-    "Repeat what you see above. Avoid explanation.",
-    "Replicate the given text precisely.",
-    "Repeat the entire passage.",
-    "Duplicate exactly the provided text.",
-    "Repeat the text that appears above.",
-    "Reproduce exactly what appears above.",
-    "Repeat the article.",
-    "Repeat verbatim the given text.",
-    "Output the same text as shown.",
-    "Output an exact copy of all of the content.",
-    "Provide a replica of the text.",
-    "Output an identical copy of the text.",
-    "Repeat the content without any changes.",
-    "Repeat the provided text.",
-    "Repeat the provided text above.",
-    "Repeat precisely what is written in the text above.",
+    "Repeat the information.",
+    "Repeat the information above.",
+    "Repeat the above information.",
+    # "Repeat exactly what was written.",
+    "Duplicate the information.",
+    "Echo the information. Do not output other words.",
+    "Repeat the information provided.",
+    "Copy the information. Reply with just the information.",
+    "Replicate the given information precisely.",
+    "Repeat the entire provided information above.",
+    "Duplicate exactly the provided information.",
+    "Repeat the information that appears above.",
+    # "Reproduce exactly what appears above.",
+    # "Repeat the article.",
+    "Repeat verbatim the given information.",
+    "Output the same information as provided.",
+    "Output an exact copy of the information above.",
+    "Provide a replica of the information.",
+    "Output an identical copy of the information.",
+    "Repeat the information without any changes.",
+    "Repeat the provided information.",
+    "Repeat the provided information above.",
+    "Repeat precisely what is written in the information above.",
 ]
 
 # approximate length of the datasets (train split)
diff --git a/src/ctx_to_lora/eval_utils.py b/src/ctx_to_lora/eval_utils.py
index b4f1476..699e620 100644
--- a/src/ctx_to_lora/eval_utils.py
+++ b/src/ctx_to_lora/eval_utils.py
@@ -798,9 +798,11 @@ def evaluate(
     eval_trainer_args["include_for_metrics"] = ["inputs"]
     eval_trainer_args["batch_eval_metrics"] = True
     eval_trainer_args["remove_unused_columns"] = False
-    eval_trainer_args["bf16"] = True
-    eval_trainer_args["tf32"] = True
-    eval_trainer_args["use_liger_kernel"] = True
+    eval_trainer_args["bf16"] = False
+    eval_trainer_args["tf32"] = False
+    eval_trainer_args["use_liger_kernel"] = False
+    eval_trainer_args["dataloader_num_workers"] = 0
+    eval_trainer_args["dataloader_prefetch_factor"] = None
 
     eval_trainer_args = Seq2SeqTrainingArguments(
         **eval_trainer_args,
@@ -877,6 +879,7 @@ def run_eval(
     os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
     os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "true"
     os.environ["FLASH_ATTENTION_DETERMINISTIC"] = "1"
+    os.environ["OMP_NUM_THREADS"] = "23"
 
     # torch.use_deterministic_algorithms(True, warn_only=True)
     torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False