| language: | |
| - en | |
| pipeline_tag: text-generation | |
| license: other | |
| license_name: llama3 | |
| license_link: LICENSE | |
| base_model: meta-llama/Meta-Llama-3-8B-Instruct | |
| tags: | |
| - causal-lm | |
| - llama-3 | |
| datasets: | |
| - athirdpath/DPO_Pairs-Roleplay-Alpaca-NSFW | |
| - allenai/UNcommonsense | |
| - ClericalAid/roleplay-scripts | |
| - fnlp/character-llm-data | |
| - IlyaGusev/pippa_scored | |
| # Nimue 8B | |
| There is a new training script for this release. | |
| The responses are shorter in the "improved" datasets. | |
| ## Prompt format | |
| The model was trained on a *zero-shot* Alpaca instruction format: | |
| ``` | |
| Below is an instruction that describes a task. Write a response that appropriately completes the request. | |
| ### Instruction: | |
| {system prompt} | |
| ### Input: | |
| User: Wait a minute. | |
| Assistant: Assistant's heart skipped a beat, she hadn't expected to meet anyone today. | |
| User: Hey, didn't I see you at the library yesterday? | |
| Traits: Shy | |
| Length: Short | |
| ### Response: | |
| ``` | |
| After several attempts, I have decided not to support multi-turn conversation for the time being. You can use labels (traits, length) to control the assistant's behavior before the response field. | |
| ## Datasets | |
| Datasets about unexpected events: | |
| - allenai/UNcommonsense (conversation format) | |
| - grimulkan/theory-of-mind (summarization) | |
| - twodgirl/tama (a cat talks to its owner) | |
| Datasets about personality traits: | |
| - allenai/soda | |
| - IlyaGusev/pippa_scored | |
| - twodgirl/ewheel | |
| - twodgirl/pi (conversation made up by Pi, the emotionally intelligent chatbot) | |
| Datasets by response length: | |
| - athirdpath/Roleplay-Alpaca-NSFW (long) | |
| - fnlp/character-llm-data (short) | |
| - twodgirl/kimiko_v3 (short) | |
| - twodgirl/theory-of-mind (short summarization) | |
| - twodgirl/pi (short) | |
| ## Personality traits | |
| There are more than 100 of them in the datasets. | |
| Affectionate, Afraid, Aggressive, Alarmed, Alert, Ambitious, Amiable, Amorous, Amused, Angry, Annoyed, Anxious, Apathetic, Apologetic, Argumentative, Aroused, Arrogant, Ashamed, Assertive, Astonished, Attentive, Bellicosity, Bitter, Bluntness, Bored, Calm, Capriciousness, Caring, Cautious, Compassionate, Competitive, Concerned, Confident, Confused, Content, Courageous, Creative, Critical, Cruelty, Curious, Defiant, Depressed, Desperate, Despondent, Determined, Disappointed, Disgusted, Disobedient, Dissatisfied, Doubtful, Efficient, Embarrassed, Empathetic, Encouraging, Enthusiastic, Envious, Excited, Exhausted, Expectant, Fidelity, Forgetful, Forgiving, Fragility, Friendly, Frugal, Frustrated, Generous, Grateful, Guilty, Happy, Hateful, Helpful, Helpless, Hesitant, Homesick, Honest, Hopeful, Hostile, Impatient, Impulsive, Indecisive, Indignant, Insecure, Insulted, Integrity, Interested, Jealous, Joyous, Kind, Kindness, Loathing, Longing, Loquacity, Lost, Loving, Loyal, Lusting, Miserable, Motivated, Nervous, Nostalgic, Optimistic, Organized, Passionate, Patient, Pensive, Persistent, Persuasive, Playful, Pleased, Polite, Protective, Proud, Rebellious, Relaxed, Relieved, Remorseful, Resilient, Restless, Reverent, Sad, Scared, Self-critical, Selfish, Sentimental, Serene, Serious, Shy, Shyness, Sleepy, Startled, Stubbornness, Superior, Supportive, Suspicious, Sympathetic, Tender, Tense, Thoughtful, Tired, Understanding, Upset, Wisdom, Worried. | |
| ## References | |
| Scherer KR. What are emotions? And how can they be measured? | |
| MIT An Affective Model of Interplay Between Emotions and Learning | |
| Scherer KR. The GRID meets the wheel | |
| Manshad Abbasi Mohsin Summarizing Emotions from Text Using Plutchik’s Wheel of Emotions | |