WIP

3 files changed, 0 insertions, 121 deletions

diff --git a/format-data.py b/format-data.py
deleted file mode 100644
index 60c13d7..0000000
--- a/format-data.py
+++ /dev/null
@@ -1,46 +0,0 @@
-import json
-
-from glob import glob
-
-from PIL import Image
-
-
-def format_example_for_gemma3_preferences(data, target_score, translation_score):
-    prompt = """<bos><start_of_turn>user
-You are a professional English-German translator and also a renowned photography critic.
-
-<start_of_image>
-Write a detailed caption for this image in a single sentence. Translate the caption into German. The output needs to be JSON, the keys being 'English' and 'German' for the respective captions. Only output the JSON, nothing else.<end_of_turn>"""
-    translation = f"""<start_of_turn>model
-{data['Translation']}<end_of_turn>"""
-    target = f"""<start_of_turn>model
-{data['German']}<end_of_turn>"""
-
-    if target_score == translation_score:
-        return None, None, None
-    elif target_score > translation_score:
-        return prompt, target, translation
-    else:
-        return prompt, translation, target
-
-
-def main():
-    with open("baseline/target.gemba-gpt4.1.scores", "r") as f:
-        scores_target = [int(line.strip()) for line in f.readlines()]
-    with open("baseline/translation.gemba-gpt4.1.scores", "r") as f:
-        scores_translation = [int(line.strip()) for line in f.readlines()]
-
-    for index, filename in enumerate(sorted(glob("baseline/*.jsonl"))):
-        with open(filename, "r") as f:
-            data = json.loads(f.read())
-        prompt, c, r = format_example_for_gemma3_preferences(data, scores_target[index], scores_translation[index])
-        print(f"{c=} {scores_target[index] > scores_translation[index]}")
-
-
-    from transformers import AutoTokenizer
-    model_id = "google/gemma-3-4b-it"
-    tokenizer = AutoTokenizer.from_pretrained(model_id)
-        
-
-if __name__ == "__main__":
-    main()
diff --git a/inference2.py b/inference.py
index 67e633a..67e633a 100644
--- a/inference2.py
+++ b/inference.py
diff --git a/test.py b/test.py
deleted file mode 100644
index 35ceb5f..0000000
--- a/test.py
+++ /dev/null
@@ -1,75 +0,0 @@
-import torch
-import requests
-from PIL import Image
-from transformers import AutoProcessor, AutoModelForCausalLM
-
-from glob import glob
-import json
-import os
-from datasets import Dataset
-
-
-def make_chat_data(base="./baseline"):
-    dataset = []
-    for filename in sorted(glob(f"{base}/*.jsonl"))[0:1]:
-        with open(filename, "r") as f:
-            data = json.loads(f.read())
-        image_path = f"../Images/{os.path.basename(filename).removesuffix(".jsonl")}.jpg"
-        image = Image.open(image_path).convert("RGB")
-        chat = [{
-                "role": "user",
-                "content": [
-                    {"type": "text", "text": "You are a professional English-German translator and also a renowned photography critic.\n\nWrite a detailed caption for this image in a single sentence. Translate the caption into German. The output needs to be JSON, the keys being 'English' and 'German' for the respective captions. Only output the JSON, nothing else."},
-                    {"type": "image"}
-                ]
-                },
-                #{ "role": "assistant",
-                #  "content": [{"type": "text", "text": '{"English": ' +  json.dumps(data["English"]) + ', "German": ' + json.dumps(data["Translation"]) + '}'}]
-                #}
-        ]
-        item = {"image": image, "chat": chat}
-        dataset.append(item)
-
-    return Dataset.from_list(dataset)
-
-
-model_id = "google/gemma-3-4b-it"
-processor = AutoProcessor.from_pretrained(model_id, use_fast=True)
-model = AutoModelForCausalLM.from_pretrained(
-    model_id,
-    dtype=torch.bfloat16,
-    device_map="auto",
-    attn_implementation="eager",
-)
-device = model.device
-dataset = make_chat_data()
-chat_prompt = processor.tokenizer.apply_chat_template(
-    [item["chat"] for item in dataset],
-    tokenize=False,
-	add_generation_prompt=True,
-)
-
-print(dataset[0])
-
-inputs = processor(
-    text=chat_prompt,
-    images=[item["image"] for item in dataset],
-    return_tensors="pt"
-).to(device)
-
-print("Keys in the output:", inputs.keys())
-
-input_ids = inputs["input_ids"]
-print("\nShape of input_ids:", input_ids.shape)
-print("input_ids:", input_ids)
-decoded_text = processor.decode(input_ids[0], skip_special_tokens=False)
-print("\nDecoded input_ids (showing special tokens):")
-print(decoded_text)
-pixel_values = inputs["pixel_values"]
-print("\n--- Generating a Response ---")
-output = model.generate(
-    **inputs,
-    max_new_tokens=100
-)
-generated_text = processor.decode(output[0], skip_special_tokens=True)
-print("\nModel's response:\n", generated_text)