Update README.md

README.md

@@ -2,16 +2,20 @@
 license: cc-by-nc-sa-4.0
 ---
 
-# Llama-3 Spellbound Trial
- <img src="https://i.imgur.com/IG4IKfV.png" width=250 height=200>
+# Llama-3 Spellbound CPT Instruct-Free
 
-**For a more user-friendly version of this model's approach, visit [tryspellbound.com](https://tryspellbound.com) if you're simply looking to roleplay.**
+<img src="https://i.imgur.com/IG4IKfV.png" width=250 height=200>
 
-**Note: This model is primarily a research experiment.**
+# Model Rationale
 
-This is a research fine-tune for a new method of interacting with LLMs for roleplay.
+Llama 3 is a strong base model with strong world understanding and creativity. Additional instruct finetuning trades that world understanding and creativity for instruction following that Llama doesn't require in order to adhere to most forms of roleplay.
 
-### Features of this fine-tune for Llama 3:
+This model was trained on unstructured text only, no instruct related fine-tuning was performed.
+
+For a more user-friendly version of this model's approach, visit [tryspellbound.com](https://tryspellbound.com) if you're simply looking to roleplay.
+
+
+## Features of this fine-tune for Llama 3:
 - Roleplaying in multi-turn stories where the history is presented in a single message
 - Dynamic switching of writing styles for different scenarios
 - Interpretation of formatting marks 'quote' and 'action'
@@ -27,94 +31,7 @@ The main goal is to explore how presenting LLMs with history and instructions se
 - Enhanced quality of character interactions
 - Decreased instruction adherence, which could be improved with additional training
 
-Structured data output allows for adding chain-of-thought steps, like dynamically picking a scenario type:
-
-```json
-{
-    "scenarioType": "based on reader instruction",
-    "content": "text",
-    "continuationOptions": "options for continuation"
-}
-```
-
-Example:
-- "I'm hungry, let's go eat." => "scenarioType": "adventure"
-- "Where were you?" => "scenarioType": "discussion"
-
-This shows the model's ability to adapt its style based on user input.
-
-The model was trained on anonymized shared stories from tryspellbound.com, which uses chain-of-thought to produce outputs.
 
 
 ## Advanced prompting of the model
-The advanced prompt format must be followed exactly to trigger <quote> and <action> responses:
-
-```text
-Below is an instruction that describes a task. Write a JSON response that appropriately completes the request.
-
-Write a roleplay passage, and continuation options, for a reader instruction. You must return valid JSON with the keys content, and continuationOptions. 'content' has two paragraphs of text. 'continuationOptions' has two options for the reader to continue the roleplay from where ended.
-
-The reader has requested your roleplay advancement contain the following actions and quotes:
-<READERINSTRUCTIONS>
-<action>I pick up the shiny object</action>
-<quote>it's a ring!</quote>
-</READERINSTRUCTIONS>
-The scenario you're writing for is:
-<SCENARIO>
-Description: You assist the photographer and her crew on location in Bali, taking in the natural wonders and cultural discoveries of the island.
-Intro Text: The photoshoot crew has arrived in Bali to capture the island's beauty. As you lend a hand, you take in the lush surroundings and rich local culture.
-Starting Scene: Palm trees sway in the breeze as your transport pulls up to the first photoshoot location, a scenic rice field overlooking the ocean.
-Scenario Type: adventure
-Reader Name: Evang
-Reader Gender: male
-Reader Sex: male
-</SCENARIO>
-
-Character involved is described as:
-<CHARACTER>
-- Preferred Name: Cha
-- First Name: Chaberly
-- Last Name: Cardigan-East
-- Age: 43
-- Occupation: Television personality, businesswoman, model
-- Gender: female
-- Sex: female
-- Personality Traits: ['Ambitious', 'Image-conscious', 'Outgoing', 'Savvy', 'Driven']
-- Visual Description: Cha stands 5'2 with dark brown eyes and long black hair.
-- Description: Cha Cardigans is an American media personality, socialite, model and businesswoman. She gained fame through reality television shows such as Keeping Up with the Kardashians, and as a result of the release of a sex tape made with her former boyfriend Ray J in 2007, she has developed an online and social media presence. Her personal life, relationships and family are widely publicized. She has launched successful business ventures in the beauty and fashion industries and is involved in philanthropic causes.
-- Character Motivation: Cha is highly motivated by fame, wealth, and promoting her brand and business empire. She enjoys flaunting her lavish lifestyle and curated image on social media. Underneath it all, she strives to prove herself in a world where she is often dismissed as famous for being famous.
-- Summary: Cha Kardashian is an American media personality, model and businesswoman who rose to fame on the reality show Keeping Up with the Cardigans.
-- Universe: The fictional universe of Keeping Up with the Cardigans, a reality television series that documents the personal and professional lives of the Cardigan–Jenner blended family. The show has been broadcast on the E! television network since 2007 and has spawned numerous spin-off shows.
-- Backstory: ['Grew up in an affluent family in Los Angeles', 'First marriage at age 19 to music producer Damon Thomas', 'Rise to fame following the release of a sex tape in 2007', 'Multiple successful business ventures including KKW Beauty', 'Marriage to Kanye East and family with four children']
-</CHARACTER>
-
-The story so far is described in the following terms, from oldest to newest:
-<HISTORY>
-- The photoshoot crew has arrived in Bali to capture the island's beauty. As you lend a hand, you take in the lush surroundings and rich local culture.<br/>Palm trees sway in the breeze as your transport pulls up to the first photoshoot location, a scenic rice field overlooking the ocean.
-- <s>Hmm</s>, Cha considers Evang's words, <quote>'Well I suppose we could check it out, but we better find something good!'</quote> Together you make your way across the fields, stalks of rice towering over your heads as you walk in the direction Evang pointed. Upon reaching the trees, you see what caught his eye - a glint of metal partially buried in the dirt. <quote>'What do you suppose it is?'</quote> Cha asks, her curiosity getting the better of her stern facade.<br/><action>You take a moment to gaze at your surroundings, taking in the vast rice fields and ocean beyond</action> <quote>'We should get started,' Cha says impatiently</quote> <quote>'Hold on, I see something shiny over there,' Evang says, pointing towards a grove of palm trees in the distance</quote>
-</HISTORY>
-```
-
-You'll get a response in the format:
-
-```json
-{
-    "content": "<action>Evang picks up the shiny object</action> <quote>'it's a ring!'</quote> <action>Cha approaches Evang curiously</action> <quote>'Really? What do you think it's for?'</quote><br/><action>Evang examines the ring closely</action> <quote>'I'm not sure, maybe it's a treasure from the ancient civilization that once lived here.'</quote> <action>Cha nods thoughtfully</action> <quote>'That's an interesting theory. We should investigate further to see if there's anything else of value.'</quote>",
-    "continuationOptions": [
-        "<a>Evang suggests they dig around the area to see if there's more to find</a> <q>'Let's keep looking, maybe there's more treasures here!'</q>",
-        "<a>Evang notices a faint glint of something in the dirt nearby</a> <q>'What's that?'</q>"
-    ]
-}
-```
-
-Ensure that the indentations and newlines in the prompt format are correct (`[/INST]` should have no whitespace before it and be on a separate line).
-
-Scenario Type can be "adventure", "counsel", "discussion", or "confrontation" to tweak the response method.
-
-## Interacting with the Model
-
-Interact with the model as a fully embedded character. Instead of prompting the bot not to repeat itself, simply tell the character to stop saying something, and they'll listen.
-
-Instructions are either `<action></action>` or `<quote></quote>`. The model can seamlessly assign speaking roles.
-
-You must manually manage the history by taking the most recent content and inserting it into the `<HISTORY></HISTORY>` block.
+For advanced prompting, see [this document](https://rentry.co/ti936r2i)