E

EVA Llama 3.33 70B

Eva Unit 01
128K
Context
N/A
Input /1M
N/A
Output /1M
N/A
Max Output
Run API Call Open in Chat

EVA Llama 3.33 70B

Model Overview

Property Value
Model ID eva-unit-01/eva-llama-3.33-70b
Full Name EVA Llama 3.33 70B
Version v0.1
Created December 16, 2024
Creator eva-unit-01 (Kearm, Auri, and Cahvay)
Type Roleplay & Storywriting Specialist
Base Model Llama-3.3-70B-Instruct (Meta)

Description

EVA Llama 3.33 70B is a roleplay and storywriting specialist model. It is a complete parameter fine-tune of the Llama-3.3-70B-Instruct base model using a mixture of synthetic and natural data. The model incorporates the Celeste 70B 0.1 data mixture to enhance versatility, creativity, and flavor for creative writing tasks.

Key Features

  • Optimized for roleplay and creative writing
  • Improved long-context comprehension and recall
  • Reduced overfitting compared to v0.0
  • Enhanced stability and reduced repetition
  • DELLA linear merge technique for improved quality

Technical Specifications

Specification Value
Context Window 128,000 tokens
Parameters 70 billion
Context Length 16,384 tokens
Data Type BF16 (bfloat16)
Input Modalities Text
Output Modalities Text
Instruction Type llama3
Stop Sequences <|eot_id|>, <|end_of_text|>

Pricing

Note: This model is available on LangMart. Pricing information was not publicly listed at the time of documentation. Please check LangMart's model page for current pricing.

Typical LangMart pricing for 70B models ranges from $0.50-$2.00 per 1M tokens depending on the provider.

Use Cases

This model excels at:

  • Roleplay scenarios - Character-driven conversations and interactions
  • Creative writing - Stories, narratives, and fiction
  • Storywriting - Long-form creative content
  • Character development - Creating and maintaining character personas
  • Worldbuilding - Developing rich fictional settings

Providers

This model is available through:

  • LangMart - Primary API access
  • Hugging Face - Model weights available for self-hosting

Hugging Face Model

Training Data

The model was trained on a curated mixture of datasets:

  1. Celeste 70B 0.1 data mixture (minus Opus Instruct subset)
  2. Kalomaze's Opus_Instruct_25k dataset (filtered for refusals)
  3. ChatGPT-4o-WritingPrompts by Gryphe (1k rows subset)
  4. Sonnet3.5-Charcards-Roleplay by Gryphe (2k rows subset)
  5. Synthstruct and SynthRP datasets by Epiculous
  6. Dolphin-2.9.3 subset (filtered not_samantha + small systemchat subset)

v0.1 Improvements

The v0.1 release includes a DELLA (Differential Evolution Linear Algebra) linear merge of v0.0 with an unreleased checkpoint, resulting in:

  • Reduced overfitting
  • Improved long context comprehension/recall
  • Reduced repetition

Merge Configuration

models:
  - model: EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.0
    parameters:
      density: 0.6
      weight: 0.3
      lambda: 1.1
      epsilon: 0.35
  - model: EVA-UNIT-01/LLaMA-EVA-3.33-70B-v0.0
    parameters:
      density: 0.45
      weight: 0.7
      lambda: 1.1
      epsilon: 0.4

merge_method: della_linear
base_model: meta-llama/Llama-3.1-70B
parameters:
  normalize: true
  int8_mask: true
dtype: bfloat16

Sampler Configuration

{
  "temperature": 1.0,
  "min_p": 0.05,
  "repetition_penalty": 1.03
}

SillyTavern Preset

A preset is available via Virt-io: EV01-llama.json

Usage Examples

LangMart API

curl https://api.langmart.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $LANGMART_API_KEY" \
  -d '{
    "model": "eva-unit-01/eva-llama-3.33-70b",
    "messages": [
      {
        "role": "system",
        "content": "You are a creative writing assistant specializing in immersive storytelling."
      },
      {
        "role": "user",
        "content": "Write the opening scene of a fantasy adventure in a misty forest."
      }
    ],
    "temperature": 1.0,
    "max_tokens": 1024
  }'

Python (OpenAI SDK)

from openai import OpenAI

client = OpenAI(
    base_url="https://api.langmart.ai/v1",
    api_key="your-langmart-api-key"
)

response = client.chat.completions.create(
    model="eva-unit-01/eva-llama-3.33-70b",
    messages=[
        {
            "role": "system",
            "content": "You are a creative writing assistant specializing in immersive storytelling."
        },
        {
            "role": "user",
            "content": "Write the opening scene of a fantasy adventure in a misty forest."
        }
    ],
    temperature=1.0,
    max_tokens=1024,
    extra_body={
        "repetition_penalty": 1.03
    }
)

print(response.choices[0].message.content)

LangMart Gateway

curl https://api.langmart.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-your-api-key" \
  -d '{
    "model": "eva-unit-01/eva-llama-3.33-70b",
    "messages": [
      {
        "role": "system",
        "content": "You are a creative writing assistant."
      },
      {
        "role": "user",
        "content": "Write a dramatic scene."
      }
    ],
    "temperature": 1.0
  }'

Prompt Format

This model uses the Llama3 chat template:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{system_message}<|eot_id|><|start_header_id|>user<|end_header_id|>

{user_message}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

License

  • Base License: Llama 3.3 Community License Agreement (Meta)
  • Permitted Uses: Personal, research, and commercial use under L3.3 license terms
  • Restrictions: Infermatic Inc and its employees/paid associates cannot utilize, distribute, or use EVA models for any purpose
  • Policy: Subject to Acceptable Use Policy for Llama Materials

Credits & Acknowledgments

  • Model Creators: Kearm, Auri, and Cahvay
  • Dataset Contributors: Gryphe, Lemmy, Kalomaze, Nopm, Epiculous, CognitiveComputations
  • Support & Testing: Allura-org (support, feedback, beta-testing, quality control)
  • Dataset Filtering: Cahvay
  • Merge Method Reference: arXiv:2406.11617 (DELLA)

Last updated: December 23, 2024