prithivMLmods

Q3.6-27B-DS-v4-Flash-DA

README

License: apache-2.0

Key Highlights

DeepSeek V4 Flash Distillation: Fine-tuned using distilled reasoning traces derived from DeepSeek V4 Flash generations for enhanced reasoning depth and contextual understanding.
Distilled-Abliterated (DA): Applies refusal direction analysis and ablation-based strategies to reduce internal refusal behaviors while maintaining reasoning quality.
Qwen3.6 Backbone: Built on top of Qwen/Qwen3.6-27B via prithivMLmods/Qwen3.6-27B-abliterated-rMAX for strong instruction-following and reasoning performance.
Instruction + Reasoning Fusion: Handles instruction-following and complex multi-step reasoning tasks seamlessly.
High-Coherence Outputs: Maintains consistency across long generations with improved contextual grounding.
27B Scale Performance: Delivers high-capacity reasoning suitable for advanced research and complex tasks.

Datasets Used and Training Details

Table with columns: Category, Details
Category	Details
Base Model	Qwen/Qwen3.6-27B
Intermediate Base	prithivMLmods/Qwen3.6-27B-abliterated-rMAX
Final Model Size	27B Parameters
Training Type	Distillation + abliteration
Objective	Preserve reasoning quality while reducing refusal behaviors and improving instruction-following reliability
Reasoning Dataset	Jackrong/DeepSeek-V4-Distill-8000x (4000 random samples used)
Alignment / Evaluation Dataset	prithivMLmods/harm_bench
Training Pipeline	TRL (Transformer Reinforcement Learning)
Training Focus	Structured reasoning, long-chain thinking, robustness across diverse prompts

Quick Start with Transformers

bash

pip install transformers==5.8.0# or latestpip install git+https://github.com/huggingface/transformers.git

python

from transformers import Qwen3_5ForConditionalGeneration, AutoProcessorimport torch model = Qwen3_5ForConditionalGeneration.from_pretrained(    "prithivMLmods/Q3.6-27B-DS-v4-Flash-DA",    torch_dtype="auto",    device_map="auto") processor = AutoProcessor.from_pretrained(    "prithivMLmods/Q3.6-27B-DS-v4-Flash-DA") messages = [    {        "role": "user",        "content": [            {                "type": "text",                "text": "Generate a highly detailed caption of a futuristic city skyline at sunset."            }        ],    }] text = processor.apply_chat_template(    messages,    tokenize=False,    add_generation_prompt=True) inputs = processor(    text=[text],    padding=True,    return_tensors="pt").to("cuda") generated_ids = model.generate(    **inputs,    max_new_tokens=512) generated_ids_trimmed = [    out_ids[len(in_ids):]    for in_ids, out_ids in zip(inputs.input_ids, generated_ids)] output_text = processor.batch_decode(    generated_ids_trimmed,    skip_special_tokens=True,    clean_up_tokenization_spaces=False) print(output_text)

Intended Use

Reasoning & Chain-of-Thought Tasks: Deep multi-step reasoning powered by DeepSeek-distilled traces
Instruction Following: Hybrid prompts requiring both instruction adherence and reasoning
Red-Teaming & Alignment Research: Evaluating reduced-refusal systems and refusal direction analysis
Local High-Performance Deployment: Multi-GPU or optimized inference setups
Research on Abliteration: Studying the effects of ablation-based training on reasoning preservation

Limitations & Risks

Important Note: This model intentionally minimizes built-in safety refusals.

Sensitive Content Risk: May produce unrestricted or controversial outputs
User Responsibility: Requires careful and ethical usage
High Compute Demand: Requires significant VRAM or optimized quantization for efficient inference
Abliteration Trade-offs: Reduced refusal may impact safety alignment and output filtering

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

prithivMLmods

Model Tree

Base

prithivMLmods/Qwen3.6-27B-abliterated-rMAX

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality