Wilfrid28/llama3-meld-wilfrid

WILFRID — #1 8B Model on MELD (0.6607 Weighted F1)

The highest-performing 8B model ever on the MELD emotion recognition benchmark.

Results (Test Set)

Emotion	Precision	Recall	F1	Support
anger	0.6736	0.4725	0.5554	345
disgust	0.3768	0.3824	0.3796	68
fear	0.3571	0.1000	0.1562	50
joy	0.6921	0.5423	0.6081	402
neutral	0.7130	0.9156	0.8017	1256
sadness	0.6447	0.2356	0.3451	208
surprise	0.6263	0.6263	0.6263	281
accuracy			0.6847	2610
weighted F1			0.6607

Trained in ~80 minutes on a single RTX 5090 using 4-bit QLoRA.

Quick Inference Example

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("Wilfrid28/llama3-meld-wilfrid", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Wilfrid28/llama3-meld-wilfrid")

prompt = """Conversation:
Ross: We were on a break!
Rachel: No we weren't!

Current turn:
Ross: WE WERE ON A BREAK!!
Emotion (one word):"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=5, temperature=0.1)
emotion = tokenizer.decode(output[0], skip_special_tokens=True).split()[-1]
print(emotion)  # → anger


### Live Demo
**Try the Emotion Oracle Dashboard**: https://huggingface.co/spaces/Wilfrid28/wilfrid-oracle

The most beautiful emotion detection interface ever built.