Triangle104
commited on
Update README.md
Browse files
README.md
CHANGED
@@ -19,6 +19,114 @@ tags:
|
|
19 |
This model was converted to GGUF format from [`prithivMLmods/Phi-4-Empathetic`](https://huggingface.co/prithivMLmods/Phi-4-Empathetic) using llama.cpp via the ggml.ai's [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space.
|
20 |
Refer to the [original model card](https://huggingface.co/prithivMLmods/Phi-4-Empathetic) for more details on the model.
|
21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
## Use with llama.cpp
|
23 |
Install llama.cpp through brew (works on Mac and Linux)
|
24 |
|
|
|
19 |
This model was converted to GGUF format from [`prithivMLmods/Phi-4-Empathetic`](https://huggingface.co/prithivMLmods/Phi-4-Empathetic) using llama.cpp via the ggml.ai's [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space.
|
20 |
Refer to the [original model card](https://huggingface.co/prithivMLmods/Phi-4-Empathetic) for more details on the model.
|
21 |
|
22 |
+
---
|
23 |
+
Model details:
|
24 |
+
-
|
25 |
+
[Phi-4 Empathetic finetuned] from Microsoft's Phi-4 is an advanced open model built upon a blend of high-quality synthetic datasets, data from filtered public domain websites, and carefully selected academic resources. It excels at responsible human-like reasoning, empathetic dialogue, and emotional thought generation. The model is designed to engage in nuanced, thoughtful conversations, with outputs that can include special characters and emojis for expressive communication. π
|
26 |
+
|
27 |
+
Phi-4 Empathetic employs a sophisticated safety post-training approach, leveraging both open-source and proprietary datasets. Safety alignment is achieved using a combination of SFT (Supervised Fine-Tuning) and DPO (Direct Preference Optimization), targeting responsible interaction and emotional awareness in diverse contexts.
|
28 |
+
Dataset Info
|
29 |
+
|
30 |
+
Phi-4 Empathetic is fine-tuned on a carefully curated dataset tailored for empathetic and responsible reasoning tasks. The dataset incorporates the Chain of Thought (CoT) methodology, emphasizing logical reasoning, emotional nuance, and step-by-step thought processes. Additionally, it includes data optimized for generating responses that resonate with human emotions, making it ideal for:
|
31 |
+
|
32 |
+
Emotional Support Applications π€
|
33 |
+
Responsible Conversations π¬
|
34 |
+
Thoughtful Problem-Solving π§
|
35 |
+
|
36 |
+
Run with Transformers
|
37 |
+
|
38 |
+
# pip install accelerate
|
39 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
40 |
+
import torch
|
41 |
+
|
42 |
+
tokenizer = AutoTokenizer.from_pretrained("prithivMLmods/Phi-4-Empathetic")
|
43 |
+
model = AutoModelForCausalLM.from_pretrained(
|
44 |
+
"prithivMLmods/Phi-4-Empathetic",
|
45 |
+
device_map="auto",
|
46 |
+
torch_dtype=torch.bfloat16,
|
47 |
+
)
|
48 |
+
|
49 |
+
input_text = "Can you share some words of encouragement for someone feeling down?"
|
50 |
+
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
|
51 |
+
|
52 |
+
outputs = model.generate(**input_ids, max_new_tokens=32)
|
53 |
+
print(tokenizer.decode(outputs[0]))
|
54 |
+
|
55 |
+
You can ensure correct formatting for empathetic dialogue by using tokenizer.apply_chat_template as follows:
|
56 |
+
|
57 |
+
messages = [
|
58 |
+
{"role": "user", "content": "Can you share some words of encouragement for someone feeling down?"},
|
59 |
+
]
|
60 |
+
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", return_dict=True).to("cuda")
|
61 |
+
|
62 |
+
outputs = model.generate(**input_ids, max_new_tokens=256)
|
63 |
+
print(tokenizer.decode(outputs[0]))
|
64 |
+
|
65 |
+
Intended Use
|
66 |
+
|
67 |
+
The Phi-4 Empathetic model is optimized for applications that require thoughtful and emotionally aware interactions. Below are some suggested use cases:
|
68 |
+
|
69 |
+
Emotional Support & Counseling π
|
70 |
+
Providing thoughtful responses to users seeking emotional encouragement or advice.
|
71 |
+
Generating empathetic messages for mental health and well-being applications.
|
72 |
+
|
73 |
+
Responsible Dialogue Generation π£οΈ
|
74 |
+
Engaging in nuanced conversations with a focus on fairness, safety, and ethical considerations.
|
75 |
+
Ensuring that interactions remain respectful and aligned with safety guidelines.
|
76 |
+
|
77 |
+
Creative Writing Assistance βοΈ
|
78 |
+
Helping users craft emotionally engaging content, including stories, poems, and personal messages.
|
79 |
+
Assisting in generating content enriched with special characters and emojis for expressive communication.
|
80 |
+
|
81 |
+
Educational Tools π
|
82 |
+
Offering step-by-step explanations with an empathetic tone for better understanding.
|
83 |
+
Generating thoughtful Q&A responses for various subjects.
|
84 |
+
|
85 |
+
Customer Support π€
|
86 |
+
Automating empathetic responses to customer queries.
|
87 |
+
Handling emotionally sensitive customer service interactions with care.
|
88 |
+
|
89 |
+
Social Media Engagement π±
|
90 |
+
Generating creative, engaging, and emotionally resonant posts for social media platforms.
|
91 |
+
Providing personalized message suggestions enriched with emojis and special characters.
|
92 |
+
|
93 |
+
Limitations
|
94 |
+
|
95 |
+
While Phi-4 Empathetic is highly capable, it has certain limitations users should be aware of:
|
96 |
+
|
97 |
+
Bias and Fairness:
|
98 |
+
Despite extensive safety alignment, biases may still emerge in the modelβs responses. Users should exercise discretion, particularly in sensitive contexts.
|
99 |
+
|
100 |
+
Emotional Nuance:
|
101 |
+
The model may occasionally misinterpret the emotional tone of a prompt, leading to less relevant or inappropriate responses.
|
102 |
+
|
103 |
+
Real-Time Knowledge:
|
104 |
+
The model's knowledge is based on the data it was trained on and does not include real-time or post-training updates. It may not reflect recent events or changes in knowledge.
|
105 |
+
|
106 |
+
Safety and Harmlessness:
|
107 |
+
Although the model is aligned with safety standards, there may still be cases where outputs require human oversight to ensure appropriateness.
|
108 |
+
|
109 |
+
Resource Requirements:
|
110 |
+
Running the model efficiently may require significant computational resources, especially in large-scale or real-time applications.
|
111 |
+
|
112 |
+
Ethical Considerations:
|
113 |
+
The model must be used responsibly, avoiding any malicious applications such as generating harmful content or spreading misinformation.
|
114 |
+
|
115 |
+
Domain-Specific Limitations:
|
116 |
+
While it performs well in general-purpose tasks, it may need further fine-tuning for highly specialized domains, such as legal, medical, or financial applications.
|
117 |
+
|
118 |
+
Special Features
|
119 |
+
|
120 |
+
Emojis & Special Characters ππ‘
|
121 |
+
The model can generate responses with emojis and special characters for expressive communication, making it ideal for social media and personal messaging applications.
|
122 |
+
|
123 |
+
Human-Like Reasoning π§
|
124 |
+
Fine-tuned for responsible reasoning and empathetic dialogue, it excels at generating thoughtful and human-like responses.
|
125 |
+
|
126 |
+
Advanced Safety Alignment π
|
127 |
+
The model employs iterative SFT and DPO techniques to ensure that its outputs are helpful, harmless, and aligned with ethical standards.
|
128 |
+
|
129 |
+
---
|
130 |
## Use with llama.cpp
|
131 |
Install llama.cpp through brew (works on Mac and Linux)
|
132 |
|