setup sft model

README.md

@@ -1,5 +1,5 @@
 ---
-title: Chinese Lantern Riddles LLM Solver V2
+title: Chinese Lantern Riddles LLM Solver V2 with SFT
 emoji: 🏮
 colorFrom: gray
 colorTo: indigo

app.py

@@ -2,6 +2,7 @@
 from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
 from langchain.vectorstores import Chroma
 from langchain.embeddings import HuggingFaceEmbeddings
+from peft import AutoPeftModelForCausalLM
 
 import gradio as gr
 import hanzidentifier
@@ -17,15 +18,14 @@ current_path=str(pathlib.Path(__file__).parent.resolve())
 llm_model_name="Qwen/Qwen1.5-0.5B-Chat"
 
 #pipe = pipeline("text2text-generation", model=model)
-model = AutoModelForCausalLM.from_pretrained(
-    llm_model_name
-)
-
-#model = AutoPeftModelForCausalLM.from_pretrained(
-#    "Qwen1.5_0.5B_Chat_sft_full/checkpoint-300",
-#    low_cpu_mem_usage=True,
+#model = AutoModelForCausalLM.from_pretrained(
+#    llm_model_name
 #)
 
+model = AutoPeftModelForCausalLM.from_pretrained(
+    "ytyeung/Qwen1.5-0.5B-Chat-SFT-riddles",
+)
+
 tokenizer = AutoTokenizer.from_pretrained(llm_model_name)
 
 # %%

requirements.txt

@@ -5,4 +5,5 @@ chromadb
 sentence-transformers
 sentencepiece
 hanzidentifier
-chinese-converter
+chinese-converter
+peft