o

generative_inference.py

@@ -35,9 +35,12 @@ MODEL_OPTIONS = {
     "GPT3.5": {"base": "openai/gpt-3.5-turbo",
                "model_id": "ft:gpt-3.5-turbo-0125:asma:gpt-3-5-turbo-absa:Bb6gmwkE"},
     "GPT4o": {"base": "openai/gpt-4o",
-              "model_id": "ft:gpt-4o-mini-2024-07-18:asma:gpt4-finetune-absa:BazoEjnp"}
+              "model_id": "ft:gpt-4o-mini-2024-07-18:asma:gpt4-finetune-absa:BazoEjnp"},
+    "DeepSeek": { "base": "deepseek-ai/deepseek-llm-7b-chat", "adapter": "deepseek7bchat-json-lora-gptparam" }
+
 }
 
+
 cached_models = {}
 
 # ✅ Reusable for both mT5 + mBART
@@ -106,3 +109,68 @@ def infer_gpt_absa(text, model_key):
     except Exception as e:
         return {"error": str(e)}
 
+
+
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+def load_deepseek():
+    base_model = AutoModelForCausalLM.from_pretrained(
+        MODEL_OPTIONS["DeepSeek"]["base"],
+        device_map="auto",
+        torch_dtype=torch.float16,
+        trust_remote_code=True
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        MODEL_OPTIONS["DeepSeek"]["adapter"],
+        trust_remote_code=True
+    )
+    model = PeftModel.from_pretrained(base_model, MODEL_OPTIONS["DeepSeek"]["adapter"])
+
+    cached_models["DeepSeek"] = (tokenizer, model)
+    return tokenizer, model
+
+
+
+def build_deepseek_prompt(review_text, output=""):
+    return f"""<|system|>
+You are an advanced AI model specialized in extracting aspects and determining their sentiment polarity from customer reviews.
+
+
+Instructions:
+1. Extract only the aspects (nouns) mentioned in the review.
+2. Assign a sentiment to each aspect: "positive", "negative", or "neutral".
+3. Return aspects in the same language as they appear.
+4. An aspect must be a noun that refers to a specific item or service the user described.
+5. Ignore adjectives, general ideas, and vague topics.
+6. Do NOT translate, explain, or add extra text.
+7. The output must be just a valid JSON list with 'aspect' and 'sentiment'. Start with `[` and stop at `]`.
+8. Do NOT output the instructions, review, or any text — only one output JSON list.
+9. Just one output and one review.
+<|user|>
+{review_text}
+<|assistant|>
+{output}"""  # ✅ include the output here
+
+
+
+def infer_deepseek(text):
+    tokenizer, model = cached_models.get("DeepSeek") or load_deepseek()
+
+    prompt = build_deepseek_prompt(text)
+    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512).to(model.device)
+
+    with torch.no_grad():
+        outputs = model.generate(
+            **inputs,
+            max_new_tokens=128,
+            do_sample=False,
+            temperature=0.0,
+            pad_token_id=tokenizer.eos_token_id
+        )
+
+    decoded = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()
+    try:
+        parsed = json.loads(decoded)
+        return parsed
+    except Exception as e:
+        return {"error": str(e), "raw": decoded}

inference.py

@@ -3,7 +3,7 @@ import json
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModel, AutoConfig
 from peft import LoraConfig, get_peft_model, PeftModel
 from modeling_bilstm_crf import BERT_BiLSTM_CRF
-from generative_inference import infer_t5_bart, infer_gpt_absa
+from generative_inference import infer_t5_bart, infer_gpt_absa, infer_deepseek
 from huggingface_hub import hf_hub_download
 
 # Define supported models and their adapter IDs
@@ -23,7 +23,8 @@ MODEL_OPTIONS = {
     "GPT3.5": {"base": "openai/gpt-3.5-turbo",
                "model_id": "ft:gpt-3.5-turbo-0125:asma:gpt-3-5-turbo-absa:Bb6gmwkE"},
     "GPT4o": {"base": "openai/gpt-4o",
-              "model_id": "ft:gpt-4o-mini-2024-07-18:asma:gpt4-finetune-absa:BazoEjnp"}
+              "model_id": "ft:gpt-4o-mini-2024-07-18:asma:gpt4-finetune-absa:BazoEjnp"},
+    "DeepSeek": { "base": "deepseek-ai/deepseek-llm-7b-chat", "adapter": "deepseek7bchat-json-lora-gptparam" }
 }
 
 
@@ -139,4 +140,7 @@ def predict_absa(text, model_choice):
     elif model_choice == 'GPT3.5' or model_choice == 'GPT4o':
         decoded = infer_gpt_absa(text, model_choice)
 
+    elif model_choice == "DeepSeek":
+        return infer_deepseek(text)
+
     return decoded