LLM EXPLORER 59,420 MODELS INDEXED

Dolphin 2.9.1 Phi 3 Kensho 4.5B Abliterated V3 by cognitivecomputations

By cognitivecomputations · 12 downloads
HF Repo ↗ Copy URL

Dolphin 2.9.1 Phi 3 Kensho 4.5B Abliterated V3 is an open-source language model by cognitivecomputations. Features: 4.5b LLM, VRAM: 13.1GB, Context: 4K, License: mit, Instruction-Based, LLM Explorer Score: 0.12.

  Autotrain compatible Base model:finetune:unsloth/ph... Base model:unsloth/phi-3-mini-... Dataset:cognitivecomputations/... Dataset:cognitivecomputations/... Dataset:cognitivecomputations/...   Dataset:internlm/agent-flan Dataset:locutusque/function-ca... Dataset:m-a-p/codefeedback-fil... Dataset:microsoft/orca-math-wo...   Dataset:teknium/openhermes-2.5   En   Endpoints compatible   Instruct   Mistral   Region:us   Safetensors   Sharded   Tensorflow
Model Card on HF 🤗: https://huggingface.co/cognitivecomputations/Dolphin-2.9.1-Phi-3-Kensho-4.5B-abliterated-v3 

Dolphin 2.9.1 Phi 3 Kensho 4.5B Abliterated V3 Parameters and Internals

Model Type 
Instruction, Conversational, Coding
Use Cases 
Areas:
Research, Commercial Applications
Primary Use Cases:
Instruction skills, Conversational skills, Coding skills, Initial agentic abilities, Function calling
Limitations:
Removed alignment and bias, Requires own alignment layer for safe deployment
Considerations:
Enjoy responsibly.
Additional Notes 
Model weights adjusted to ablate and inhibit refusals. Supports use of both merged model and adapter method for replication.
Supported Languages 
en (Proficient)
Training Details 
Data Sources:
Cognitive Computations, teknium/OpenHermes-2.5, m-a-p/CodeFeedback-Filtered-Instruction, cognitivecomputations/dolphin-coder, cognitivecomputations/samantha-data, microsoft/orca-math-word-problems-200k, Locutusque/function-calling-chatml, internlm/Agent-FLAN
Methodology:
PEFT layer replication, qLoRA fine-tuning, ablation to inhibit refusals based on 'Refusal in LLMs is mediated by a single direction'.
Context Length:
4000
Training Time:
2.5 days
Hardware Used:
8xL40S node from Crusoe Cloud
Model Architecture:
Based on Phi-3-Mini-Instruct-4k
Input Output 
Input Format:
ChatML prompt template
LLM NameDolphin 2.9.1 Phi 3 Kensho 4.5B Abliterated V3
Repository 🤗https://huggingface.co/cognitivecomputations/Dolphin-2.9.1-Phi-3-Kensho-4.5B-abliterated-v3 
Base Model(s)  Phi 3 Mini 4K Instruct   unsloth/Phi-3-mini-4k-instruct
Model Size4.5b
Required VRAM13.1 GB
Updated2025-07-13
Maintainercognitivecomputations
Model Typemistral
Instruction-BasedYes
Model Files  5.0 GB: 1-of-3   5.0 GB: 2-of-3   3.1 GB: 3-of-3
Supported Languagesen
Model ArchitectureMistralForCausalLM
Licensemit
Context Length4096
Model Max Length4096
Transformers Version4.40.2
Vocabulary Size32064
Torch Data Typebfloat16

Best Alternatives to Dolphin 2.9.1 Phi 3 Kensho 4.5B Abliterated V3

Best Alternatives
Context / RAM
Downloads
Likes
...olphin 2.9.1 Phi 3 Kensho 4.5B4K / 7.6 GB1632
...i 3 Kensho 4.5B Abliterated V34K / 13.1 GB6310
...olphin 2.9.1 Phi 3 Kensho 4.5B4K / 7.6 GB4531
...in 2.9.1 Phi 3 Kensho 4.5B AWQ4K / 3.7 GB70
Note: green Score (e.g. "73.2") means that the model is better than cognitivecomputations/Dolphin-2.9.1-Phi-3-Kensho-4.5B-abliterated-v3.