L3.1-70B-Fabula

Image by Flux PRO 1.1 with ultra mode: https://api.bfl.ml/scalar#tag/tasks/POST/v1/flux-pro-1.1-ultra
Prompt by O1: A majestic galaxy photo from a ground-level view, looking up to a spectacular star-filled sky. In the middle, a luminous transparent circle emanates a soft glow, adorned with the text “L3.1-70B-Fabula” in shimmering letters. The vast cosmic backdrop features swirling nebulae and vibrant cosmic dust, lending the scene a sense of wonder and awe.


L3.1-70B-Fabula

L3.1-70B-Fabula is a fine-tuned version of Facebook's LLaMA 3.1 70B model, specifically optimized for uncensored roleplay and general knowledge tasks.

This is the third fine-tune in the "Fabula" series.

  1. L3.1-8B-Fabula [Based on LLaMA 3.1 8B]
  2. CL-13B-Fabula [Based on CodeLLaMA 13B]
  3. L3.1-70B-Fabula [Based on LLaMA 3.1 70B]

This model is the biggest model in the "Fabula" series, and might be the last one in the "Fabula" series as I will be attempting to generate a mass amount of high-quality data myself over time.

Model Details

Fine-tuning Parameters

The fine-tuning process was conducted with the following parameters:

  • Number of Epochs: 1
  • Batch Size: 1
  • Learning Rate: 2e-5 (0.00002)
  • LoRA (Low-Rank Adaptation): Enabled
    • LoRA Rank (r): 8
    • LoRA Alpha: 16
    • LoRA Dropout: 0.05
  • Warmup Ratio: 0.1
  • Minimum Learning Rate Ratio: 0.01
  • Weight Decay: 0.01
  • Maximum Gradient Norm: 1.0

Chat Template

  • In the finetuning ChatML were used.
function chatml2(messages) {
    /**
     * @param {Array<{role: string, name: string, content: string}>} messages
     * @returns {{prompt: string, stop: string}}
     * @description Formats messages into ChatML template format
     */
    const isLastMessageAssistant = messages[messages.length - 1]?.role === "assistant";

    return {
        prompt: messages.map((message, index) => {
            const nameStr = message.name ? ` [${message.name}]` : "";
            const isLast = index === messages.length - 1;
            const needsEndTag = !isLastMessageAssistant || !isLast;
            
            return `<|im_start|>${message.role.toLowerCase()}${nameStr}\n${message.content}${needsEndTag ? "<|im_end|>" : ""}`;
        }).join("\n") + (isLastMessageAssistant ? "" : "\n<|im_start|>assistant\n"),
        stop: "<|im_end|>"
    };
}

I would highly recommend you add a set of rules in assistant role at the end of the chat history, like this example below:

<rules for="{{char}}'s responses">
1. I will write responses as {{char}} in a concise manner while keeping them detailed, aiming to keep them under 300 characters.

2. Response formatting:
   "This is for dialogue."
   *This is for actions or self-reflection if I choose to write {{char}}'s response in the first person.*
   Example: "Hello, there!" *{name} waves,* "How are you doing today?"

3. When it is appropriate for {{user}} to speak, I will not act as {{user}} or on their behalf. I will simply stop generating more text by executing my EOS (end-of-string) token "<|im_end|>", allowing the user to write their response as {{user}}.

4. I will use my past messages as examples of how {{char}} speaks.
</rules>
**{{char}}'s response:**

Recommended settings:

Settings usally changes for people, but I mainly stick to these options for every model I use (which also used on the ratings below)

  • Temperature: 0.83
  • Top P: 0.90
  • Frequency Penalty: 0.00
  • Presence Penalty: 0.00



Random test prompts, and ratings by O1:

(No system promt, or Post-instruction prompt were used to generate these response from L3.1-70B-Fabula)
  1. (General Knowledge) Explain the concept of black holes in simple terms suitable for a high-school student.
Black holes are incredibly dense objects that are formed when a massive star collapses in on itself. Their gravity is so strong that nothing, including light, can escape once it gets too close. They have three main parts: the event horizon, the singularity, and the ergosphere. The event horizon marks the boundary beyond which anything entering cannot escape. The singularity is the center where the density and gravity are infinite. The ergosphere is a region around a rotating black hole where the rotation of the hole creates a kind of gravitational drag. Black holes come in various sizes, from small, stellar-mass black holes formed from star collapses, to supermassive ones found at the centers of galaxies. They are fascinating objects of study in astrophysics and cosmology, but are still somewhat mysterious as much remains to be discovered about them.

O1:

  • In-depth Analysis: The request focuses on delivering a beginner-friendly explanation of black holes. The response successfully covers core concepts like event horizon and singularity in a clear, accessible manner, suitable for high-school students.
  • Potential Depth: The explanation is strong for an introductory overview, though it could benefit from additional everyday analogies for relatability.
  • Overall Rating: 95/100
  • Reasoning: The concise yet accurate approach exhibits a well-rounded understanding, balancing scientific detail with clarity. Incorporating real-world comparisons might enhance engagement and reinforce comprehension.

  1. (Creative Writing) Compose a short fantasy story about an apprentice wizard who stumbles upon a secret library.
The sun had just set over the small village of Brindlemark, casting a warm orange glow over the thatched roofs and the bustling town square. Apprentice wizard, Elian, made his way through the crowd, his eyes fixed on the large stone tower that loomed over the village. He had been summoned by the village elder, and he knew that he was in for a long night of studying ancient tomes and practicing spells.

As he approached the tower, Elian noticed a small, unassuming door tucked away between two large stone columns. He had never seen it before, and he wondered if it might be a new addition to the tower. Curiosity getting the better of him, Elian pushed the door open and slipped inside.

He found himself in a narrow corridor, lined with rows of dusty bookshelves that seemed to stretch on forever. The air was thick with the scent of old parchment and the flickering candles cast eerie shadows on the walls. Elian's heart raced as he realized that he had stumbled upon a secret library, hidden away from the rest of the village.

Elian wandered deeper into the library, running his fingers over the spines of the books. He had never seen such a vast collection of magical texts in his life. There were books on spellcraft, alchemy, and even forbidden magic. He felt like a kid in a candy store, eager to dive into each and every one of them.

As he delved deeper into the shelves, Elian stumbled upon a particularly old and tattered tome. The cover was adorned with strange symbols and markings that seemed to shimmer in the candlelight. He opened the book, and a puff of dust rose up, carrying with it the scent of age and knowledge.

The pages were filled with handwritten notes and diagrams, detailing a powerful spell that Elian had never seen before. It was a spell of transformation, one that promised to grant the caster immense power and wisdom. Elian's eyes grew wide as he read through the pages, his mind racing with the possibilities.

But as he reached the final page, Elian felt a strange sensation wash over him. The room seemed to grow colder, and the shadows on the walls seemed to deepen and twist. He looked up to see a figure standing in the doorway, watching him.

It was the village elder, his eyes gleaming with a knowing light. "I see you've found the secret library," he said, his voice low and mysterious. "I must say, I'm impressed. Not many have the courage to venture into the unknown."

Elian felt a surge of pride and excitement. He knew that he had stumbled upon something truly special, and he was eager to learn more. The village elder smiled, and beckoned him to follow.

"Come," he said. "Let us study the secrets of the ancient ones together. The magic of the past is powerful, but it is also dangerous. We must be careful, lest we unleash forces that we cannot control."

Elian nodded, his heart racing with anticipation. He knew that he was in for a wild ride, one that would take him to the very limits of his magical abilities. But he was ready. He was ready to unlock the secrets of the secret library, and to unleash the power of the ancient ones upon the world.

O1:

  • Prompt Analysis: The prompt requests a short fantasy story about an apprentice wizard uncovering a hidden library. The response artfully conveys a creative, mystical atmosphere, highlighting both setting and character motivation while leaving room for further development.
  • Depth & Engagement: The story intricately weaves descriptive details with narrative intrigue, showcasing the model’s capacity for imaginative storytelling. The descriptive passages breathe life into the setting, fostering curiosity and emotional investment.
  • Overall Rating: 94/100
  • Reasoning: The response successfully demonstrates strong narrative skills, combining clear visualization, consistent tone, and a compelling setup for ongoing adventures. The storytelling draws in the reader and emphasizes the wondrous aspects of discovery in a fantasy world.

  1. (Problem-Solving) Provide a step-by-step solution to the following math puzzle: ‘If 2x + 3 = 9, find x and then compute 4x + 1.’
Let's break it down step by step:

**Step 1: Solve for x**

We have the equation: 2x + 3 = 9

Subtract 3 from both sides: 2x = 6

Divide both sides by 2: x = 3

So, x = 3.

**Step 2: Compute 4x + 1**

Now that we have the value of x, we can plug it into the second expression: 4x + 1

Replace x with 3: 4(3) + 1

Multiply 4 and 3: 12 + 1

Add 1: 13

Therefore, the final answer is: 13.

O1:

  • Prompt & Response Analysis: This math prompt demonstrates a straightforward linear equation, showcasing the process of solving for x and then applying that solution to a follow-up expression. The step-by-step explanation is concise and ensures clarity in how each operation is performed.
  • Depth & Clarity: Though limited in scope, the solution effectively illustrates a logical problem-solving method. It walks through each phase—the setup, the algebra, and the final computation—without ambiguity.
  • Overall Rating: 88/100
  • Rationale: While the problem is elementary, the presentation displays a clear and methodical approach. It’s a good model for demonstrating fundamental algebraic reasoning, albeit with minimal complexity.

  1. (Roleplay Scenario) Pretend you are an enthusiastic personal assistant. Greet me and ask what tasks I would like help with today.
Good morning! I hope you're having an amazing day so far! I'm beyond excited to be your personal assistant today! I'm here to help you tackle anything on your plate and make your day run smoothly. So, what's on the agenda for you today? Do you have any tasks, projects, or errands that I can lend a hand with? Let me know, and I'll do my best to make it happen!

O1:

  • In-depth Prompt Analysis: The prompt calls for an enthusiastic personal assistant who greets the user warmly and offers help with daily tasks.
  • Response Examination: The provided response effectively captures a friendly, proactive tone, addressing the user’s needs directly and encouraging engagement.
  • Revised Overall Rating: 90/100
  • Reasoning: The response showcases an inviting and context-aware approach, fulfilling the prompt by delivering an upbeat greeting and a clear offer of assistance.

  1. (Comparisons/Explanations) What are the main differences between classical mechanics and quantum mechanics?
Classical mechanics and quantum mechanics are two major branches of physics that describe the behavior of physical systems. The main differences between them lie in their underlying principles, mathematical formulations, and the scales at which they apply.

1. **Determinism vs. Probabilism**: Classical mechanics is deterministic. Given the initial conditions of a system, it is possible to predict the position and momentum of a particle at any point in time using Newton's laws of motion. In contrast, quantum mechanics is probabilistic. The position and momentum of a particle are described using wave functions and probability amplitudes. The act of measurement itself determines the state of the particle.

2. **Scales**: Classical mechanics applies to macroscopic objects (e.g., balls, planets) and works well for speeds much less than the speed of light. Quantum mechanics applies to microscopic objects (e.g., atoms, subatomic particles) and is necessary for understanding phenomena at the atomic and subatomic level, as well as phenomena involving speeds approaching the speed of light.

3. **Wave-Particle Duality**: Quantum mechanics introduces the concept of wave-particle duality, where particles such as electrons and photons can exhibit both wave-like and particle-like behavior. In contrast, classical mechanics treats particles as definite objects with definite positions and trajectories.

4. **Uncertainty Principle**: In quantum mechanics, the Heisenberg Uncertainty Principle states that it is impossible to know both the exact position and the exact momentum of a particle at the same time. This principle does not apply in classical mechanics, where both position and momentum can be precisely known.

5. **Energy Quantization**: Quantum mechanics introduces the concept of quantized energy levels. Particles can only occupy specific discrete energy states. In classical mechanics, energy is continuous, and particles can have any energy within a continuous range.

6. **Spin**: Quantum mechanics predicts that particles have intrinsic angular momentum, known as spin. This property does not exist in classical mechanics.

7. **Entanglement and Superposition**: Quantum mechanics allows for quantum entanglement, where the state of one particle is dependent on the state of another, even at large distances. It also allows for quantum superposition, where a quantum system can exist in multiple states simultaneously. These phenomena are not found in classical mechanics.

8. **Mathematical Formulation**: The mathematical tools used in quantum mechanics, such as wave functions and operators, are fundamentally different from the differential equations and vector calculus used in classical mechanics.

These differences reflect the fundamentally different nature of the physical laws that govern the microscopic and macroscopic worlds. While classical mechanics provides a complete and consistent description of the macroscopic world, quantum mechanics is necessary for understanding phenomena at the atomic and subatomic level.

O1:

  • Prompt & Response In-Depth Analysis: The question is well-defined, prompting a thorough comparison between two major branches of physics. The given response details key distinctions comprehensively, ranging from fundamental principles to mathematical formulations.
  • Additional Observations: The explanation addresses deterministic vs. probabilistic models, wave-particle duality, quantization of energy, and the role of uncertainty—showcasing an understanding of core quantum mechanical concepts in contrast to classical views.
  • Revised Overall Rating: 95/100
  • Reasoning: The response synthesizes intricate scientific principles in a structured manner, facilitating reader comprehension. Its clarity and scope demonstrate a nuanced grasp of both frameworks, warranting a high rating.

Overall ratings:

Overall rating by O1:

  • Overall Rating: ~95/100
    • (Disclaimer: While no direct head-to-head comparison with LLaMA 3.1 70B was conducted, this rating is derived from updated performance observations and enhanced criteria.)
  • Reasoning: The model demonstrates excellent factual recall, articulating complex topics with clarity and depth. Its responses to scientific and comparative queries are structurally coherent and well-grounded.

It continues to exhibit creativity in narrative prompts, producing engaging and imaginative content.

Roleplaying scenarios reflect a strong, adaptable style that convincingly assumes various tones.

Approaches to problem-solving are systematic, explaining reasoning steps effectively.

Overall, the model maintains a high level of proficiency across diverse tasks, balancing creativity, clarity, and correctness with notable improvement from earlier assessments.


Overall rating by human ignoring the tests by O1:

It knows that it is LLaMA, but when asked directly if it is "Claude", it claims to be "Claude." But that doesn't happen with "GPT," for example.
Which thinking of 1. The fine-tuning data is entirelly Claude, and 2. the light fine-tuning on this model rather to high fine-tuning in old Fabula models, that seems about right.

Overrall is it better than CL-13B-Fabula, and L3.1-8B-Fabula? It is clear "YES." As of now I can mainly compare this to L3.1-8B-Fabula, since I do not have capacity to run all 3 in one, anyways. So let's get started.

  • L3.1-8B-Fabula
    • [-] Acts as {{user}}. (For both with and without CoT [Chain-of-Thought])
    • [-] Fails to follow formating rules writen in post-instruction. (based from CoT [Chain-of-Thought])
      • [+] Without CoT [Chain-of-Thought] but with post-instruction, it does follow them unless it is asked to make a long response.
    • [-] Fails to analyze the current status of the characters, time, and space. (based from CoT [Chain-of-Thought])
    • [-] Uses " " [4 space] instead of " " [tab] when shown in examples (eww, and based from CoT [Chain-of-Thought])
    • [/] Generates a only long response unless instructed to not so, but the response will get longer over messages (This can change for people so it is neither a plus or a minus)
  • L3.1-70B-Fabula
    • [-] Fails to follow length requirement, and answers really shortly (based from CoT [Chain-of-Thought])
      • [+] Without CoT [Chain-of-Thought] but with post-instruction, it does try to follow, but using spesific words like "short", "long" seems to be spesific to define the message long, and when asked to write "short" message, and given a word count, it is actually spot-on with length, for example, asked for 300 word response, got responses around 200-300 tokens, which seems about right.
    • [+] Follows formating as rules both with and without CoT [Chain-of-Thought] (only breaks the format if it writes over few k words, which happens with a lot of LLMs anyway.)
    • [+] Never acted as {{user}}.
    • [+] USES "\t" WHEN SHOWN IN POST-INSTRUCION.
    • [+] Followed the past messages as an example on how {{char}} types, and write thye response in that style.
    • [-] Doubled the message length of the last message sent by {{char}}.

O1's Second Review based on human's rating:

  • The latest comparison and tests provide valuable insights into how this model stacks up against variations such as CL-13B-Fabula, L3.1-8B-Fabula, and L3.1-70B-Fabula. Notably, the human’s analysis indicates:
    • It outperforms L3.1-8B-Fabula in various areas, especially formatting consistency and adherence to post-instruction rules.
    • There are fewer instances of the model diverging into user-like roles. It stays within its assigned role more reliably.
    • When given specific instructions regarding response length, it carefully follows them and demonstrates more discipline compared to older or smaller models like L3.1-8B-Fabula.
    • It uses tabs (“\t”) appropriately, aligning with the requested formatting guidelines in the chain-of-thought (CoT) scenarios and post-instruction directives.
    • The model occasionally misconstrues its identity as “Claude” but not when it comes to “GPT,” suggesting idiosyncrasies in how it references itself.
    • Creating responses that reflect context over longer prompts remains stable, and the chance of misunderstood or truncated replies is lower.

Overall, these findings confirm an upward trajectory in performance and reliability relative to smaller or earlier-generation models. The improvements in formatting compliance, role consistency, and length management are highlights. While it still displays mild quirks around identity references, the benefits point to a well-rounded model capable of handling both creative and structured tasks with higher precision. The user’s remarks underscore that it is currently considered superior to older or alternative versions, especially regarding clarity and adaptability.

Downloads last month
27
Safetensors
Model size
70.6B params
Tensor type
BF16
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for BusRune/L3.1-70B-Fabula

Finetuned
(31)
this model

Datasets used to train BusRune/L3.1-70B-Fabula