TinyKiller-NSFW-DPO-1.1B

TinyKiller NSFW DPO 1.1B

<center><img src="https://i.ibb.co/FZtCxWY/IMG-20250506-210550.jpg" alt="IMG-20250506-210550" border="0"></a></center>

---

Merge Details

TinyKiller-1.1B is a language model based on the TinyLlama-1.1B architecture, designed to deliver exceptional performance in text generation, reasoning, and programming tasks. This model has been fine-tuned using a combination of diverse and high-quality datasets, allowing it to excel across multiple domains.

---

🧠 Core Capabilities of TinyKiller-1.1B

1. Conversational Alignment and Reasoning

• Intel/orca_dpo_pairs: Provides data pairs for preference-based optimization training, improving the quality of generated responses.
• LDJnr/Verified-Camel: Offers verified conversations to strengthen coherence and accuracy in dialogue.
• HuggingFaceH4/no_robots and Doctor-Shotgun/no-robots-sharegpt: Datasets designed to prevent robotic-sounding replies, encouraging more natural, human-like interactions.

2. Toxicity Resistance

• unalignment/toxic-dpo-v0.1: Includes data contained within is "toxic"/"harmful", and contains profanity and other types of sensitive content.

3. Instruction Following and Complex Reasoning

• jondurbin/airoboros-3.2: Challenges the model with complex, multi-step tasks, enhancing its instruction-following and reasoning skills.
• LDJnr/Capybara and Doctor-Shotgun/capybara-sharegpt: Provide diverse, high-quality instruction datasets to strengthen task comprehension and execution.

4. Programming and Code Understanding

• bigcode/starcoderdata: Contains a broad collection of code across multiple languages, GitHub issues, and Jupyter notebooks, enabling effective code understanding and generation. (huggingface.co)

5. General, Deduplicated Web Data

• cerebras/SlimPajama-627B: Offers a massive, deduplicated dataset for solid and diverse language foundation training. (cerebras.ai)

6. Compact and Efficient Instruction Tuning

• OEvortex/vortex-mini: Provides instruction-tuned data optimized for small models, enhancing task efficiency and performance.

---

⚙️ Key Use Cases

• Virtual Assistants: With training from datasets like no_robots and airoboros, TinyKiller-1.1B can hold natural and coherent conversations.
• Content Moderation: Thanks to toxic-dpo-v0.1, the model can detect and manage harmful or inappropriate content.
• Code Generation & Understanding: Training with starcoderdata allows the model to assist in programming and code analysis.
• Education & Tutoring: Its ability to follow detailed instructions makes it suitable for educational applications and personalized tutoring.

---

🧩 Conclusion

TinyKiller-1.1B represents a carefully crafted integration of multiple high-quality datasets, enabling robust performance across a wide range of natural language processing tasks. Its balanced design—combining conversational ability, toxicity resistance, and code comprehension—makes it a versatile tool for many applications.

---

merge

This is a merge of pre-trained language models created using mergekit.

Merge Details

Merge Method

This model was merged using the Model Stock merge method using Doctor-Shotgun/TinyLlama-1.1B-32k-Instruct as a base.

Models Merged

The following models were included in the merge:

• Danielbrdz/Barcenas-Tiny-1.1b-DPO
• MysteriousAI/Mia-1B
• cognitivecomputations/TinyDolphin-2.8.2-1.1b-laser
• mrcuddle/tiny-darkllama-dpo

Configuration

The following YAML configuration was used to produce this model:

dtype: float32
out_dtype: bfloat16
merge_method: model_stock
base_model: Doctor-Shotgun/TinyLlama-1.1B-32k-Instruct
models:
    - model: Danielbrdz/Barcenas-Tiny-1.1b-DPO
      parameters:
        weight:
          - filter: mlp
            value: [2, 1]
          - value: 1
    - model: cognitivecomputations/TinyDolphin-2.8.2-1.1b-laser
      parameters:
        weight:
          - filter: mlp
            value: [1, 2]
          - value: 1
    - model: mrcuddle/tiny-darkllama-dpo
      parameters:
        weight:
          - filter: mlp
            value: [1, 2]
          - value: 1
    - model: MysteriousAI/Mia-1B
      parameters:
        weight:
          - filter: mlp
            value: [2, 1]
          - value: 1
    - model: Doctor-Shotgun/TinyLlama-1.1B-32k-Instruct
      parameters:
        weight:
          - filter: lm_head
            value: 0
          - value: [1, 0.5]