Upload 2 files

app.py

@@ -0,0 +1,57 @@
+import os
+import pandas as pd
+import gradio as gr
+from llama_cpp import Llama
+
+# Path to GGUF file
+MODEL_PATH = "./models/llama-3-8b-instruct.Q4_K_M.gguf"
+
+# Load quantized model with llama-cpp
+llm = Llama(
+    model_path=MODEL_PATH,
+    n_ctx=2048,
+    n_threads=8,
+    verbose=False
+)
+
+def build_prompt(source, translation):
+    return (
+        f"<|begin_of_text|><|system|>\n"
+        "You are a helpful assistant that evaluates translation quality. "
+        "Score the quality from 0 (worst) to 1 (best).\n"
+        "<|user|>\n"
+        f"Original: {source}\nTranslation: {translation}\n"
+        "How good is the translation?\n<|assistant|>"
+    )
+
+def estimate_score(source, translation):
+    prompt = build_prompt(source, translation)
+    output = llm(prompt, max_tokens=10, stop=["</s>", "\n"])
+    text = output["choices"][0]["text"].strip()
+
+    try:
+        score = float([s for s in text.split() if s.replace('.', '', 1).isdigit()][-1])
+        score = round(score, 3)
+    except:
+        score = "N/A"
+    return score
+
+def process_file(file):
+    df = pd.read_csv(file.name, sep="\t")
+    scores = []
+    for _, row in df.iterrows():
+        score = estimate_score(row["original"], row["translation"])
+        scores.append(score)
+    df["predicted_score"] = scores
+    return df
+
+demo = gr.Interface(
+    fn=process_file,
+    inputs=gr.File(label="Upload dev.tsv with 'original' and 'translation' columns"),
+    outputs=gr.Dataframe(),
+    title="LLaMA 3 8B (Quantized) MT QE",
+    description="Translation Quality Estimation using LLaMA-3-8B GGUF via llama-cpp-python on CPU"
+)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,3 @@
+llama-cpp-python==0.2.56
+pandas
+gradio