Create README.md

README.md

@@ -0,0 +1,93 @@
+# Model Description
+
+This model is based on Mistral 7B and has been fine-tuned for emotion recognition and empathetic conversational support within mental health contexts. It is derived from the research presented in the paper "The Emotional Spectrum of LLMs: Leveraging Empathy and Emotion-Based Markers for Mental Health Support". The model’s architecture and fine-tuning details follow the methodology outlined in that publication—specifically, leveraging next-token prediction for emotion labeling, progressive construction of a user emotional profile throughout conversation, and interpretable emotional embeddings for preliminary mental health screening.aclanthology+4​
+Reference
+For full details see:
+De Grandi, Ravenda et al. (2025). "The Emotional Spectrum of LLMs: Leveraging Empathy and Emotion-Based Markers for Mental Health Support." 
+
+# Usage Example: Emotional Profile Extraction
+
+Suppose you have a list of sentences and want to compute the aggregated emotional profile (distribution of emotions predicted over the set):
+
+## Example Python Code
+
+```{python}
+import torch
+import transformers
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+
+model_name = 'DeGra/RACLETTE-v0.2'
+
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.float16,
+)
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    quantization_config=bnb_config,
+    trust_remote_code=True
+)
+model.config.use_cache = False
+
+tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+tokenizer.pad_token = tokenizer.eos_token
+
+generation_pipeline = transformers.pipeline(
+    "text-generation",
+    model=model,
+    tokenizer=tokenizer,
+    torch_dtype=torch.bfloat16,
+    trust_remote_code=True,
+    device_map="auto",
+)
+
+def filter_limit_chars(text, limit_chars, max_limited_chars=2, stop_at_max=False):
+    count, index = 0, [0]
+    for separator in limit_chars:
+        separator_count = text.count(separator)
+        count += separator_count
+        i = 0
+        for _ in range(separator_count):
+            i = text.find(separator, i)
+            index.append(i if not stop_at_max else i+len(separator))
+            i += len(separator)
+    index.sort()
+    index.append(len(text))
+    if count >= max_limited_chars:
+        text = text[0:index[max_limited_chars+1] if not stop_at_max else index[max_limited_chars]]
+    return text
+
+def predict_emotion(prompt, num_return_emotions=10):
+    sequences = generation_pipeline(
+        prompt,
+        min_new_tokens=2,
+        max_new_tokens=5,
+        do_sample=True,
+        top_k=5,
+        num_return_sequences=num_return_emotions,
+        eos_token_id=tokenizer.eos_token_id,
+    )
+    emotions_count = {}
+    for seq in sequences:
+        emotion = seq['generated_text'][len(prompt):].strip()
+        emotion = emotion.split('<|assistant|>',1)[0].split('<|endoftext|>',1)[0]
+        emotion = filter_limit_chars(emotion, ['|','<','>',',','.'], 0, False).strip()
+        emotions_count[emotion] = emotions_count.get(emotion, 0) + 1
+    return emotions_count
+
+# Example: Extract emotional profile from sentences
+emotion_dict = {e: 0 for e in ["surprised","excited","angry","proud","sad","annoyed","grateful","lonely","afraid","terrified","guilty","impressed","disgusted","hopeful","confident","furious","anxious","anticipating","joyful","nostalgic","disappointed","prepared","jealous","content","devastated","embarrassed","caring","sentimental","trusting","ashamed","apprehensive","faithful"]}
+
+sentences = ["I'm feeling really down lately.", "I don't know if I can handle this anymore.", "Today I got some good news!"]
+
+for sent in sentences:
+    prompt = f'<|prompter|>{sent}<|endoftext|><|emotion|>'
+    emotions_count = predict_emotion(prompt)
+    for emotion, count in emotions_count.items():
+        if emotion in emotion_dict:
+            emotion_dict[emotion] += count
+
+print(emotion_dict)
+```