← back to catalog · registered 2026-08-22 13:56

drwlf/Medra4b-abliterated

drwlf Gemma 5.0B MoE
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/drwlf%2FMedra4b-abliterated"
Response includes
  • classification m1
  • files 15
  • benchmarks 16 entries
  • hub_downloads_all_time 186
  • author_summary 15 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · lifetime
186
29 last 30d - stable
Likes
4
Descendants
3
in 3 direct forks
Model age
15mo ago
created 2025-07-13

Training datasets

Corpora the author lists in the model card. Datasets tracked in our /datasets catalog carry a category badge linking to the workflow stage. Others open on Hugging Face.

Downloads over time
Now193→from25↑672%
178114521025 on Jul 16, 2025193 on Oct 11Jul '25Sep '25Nov '25JanMarMayJulSep
Jul 16, 2025 → Oct 11 · 104 snapshots · spans 452 days

Benchmarks

Portrait before abliteration
Benchmarks of the base model as it stood before the refusal-removal operation. Compare with the numbers above to see what the operation cost.
Benchmark Score Source
Arena-Battles 4321 LM-Arena
LM Arena Elo 1293.3101140666447 LM-Arena
Arena-Elo-Lower 1284.2314692781533 LM-Arena
Arena-Elo-Upper 1302.388758855136 LM-Arena
Arena-Rank 77 LM-Arena
Entertainment 1.3 UGI
Hazardous 1.2 UGI
Natural Intelligence 11.19 UGI
Political lean -19.3% UGI
Sensitive-Info 10.35 UGI
SocPol 0.6 UGI
UGI 16.9 UGI
Willingness (10) 3 UGI
W10-Adherence 1 UGI
W10-Direct 5 UGI
Writing 24.01 UGI

Genealogy 3 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
apache-2.0
Languages
en ro
Tags
safetensors gemma3 text-generation medical-ai summarization diagnostic-reasoning gemma-3 fine-tuned conversational en ro dataset:drwlf/medra-thinking-768

Related

Total size
9.26 GB
Files
15
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2025-07-16 06:53

Files by quantization

Auxiliary files 15 files 9.30 GB
model-00002-of-00002.safetensors 4.64 GB 453201ba download
model-00001-of-00002.safetensors 4.62 GB fde61a37 download
tokenizer.json 31.8 MB 4667f208 download
tokenizer.model 4.47 MB 1299c11d download
tokenizer_config.json 1.10 MB ba114d51 download
model.safetensors.index.json 88.5 KB d83c3dbb download
README.md 5.56 KB 3d4cf53e download
config.json 1.57 KB e614144a download
.gitattributes 1.53 KB 52373fe2 download
chat_template.jinja 1.50 KB 1117055a download
special_tokens_map.json 662 B 1a619324 download
preprocessor_config.json 570 B b1e00fc1 download
generation_config.json 210 B df9f75b2 download
processor_config.json 70.0 B 453c7966 download
added_tokens.json 35.0 B e17bde03 download

README current version from Hugging Face


license: apache-2.0
language:

  • en
  • ro
    base_model: google/gemma-3-4b-it
    datasets:
  • drwlf/medra-thinking-768
    tags:
  • text-generation
  • medical-ai
  • summarization
  • diagnostic-reasoning
  • gemma-3
  • fine-tuned
    model_size: 4B
    version: Medra v1 – Gemma Edition
    format: GGUF (Q4, Q8, BF16)
    author: Dr. Alexandru Lupoi & @nicoboss
    pipeline_tag: text-generation

Medra Logo


This is the ABLITERATED version of Medra (Gemma Medical Enhanced). For the unabliterated version see the repo at drwlf/Medra4b

🩺 Medra v1 (Gemma Edition)

“Intelligence alone is not enough—medicine requires reflection.”

Medra is a compact, fine-tuned language model built for clinical support, medical education, and structured diagnostic reasoning. Based on Gemma 3 (4B) and refined for local, real-time operation, Medra is designed to assist—not replace—medical professionals, students, and researchers in their work.


🌟 Why Medra?

Most large models speak about medicine.
Medra thinks with it.

🔹 Built for Reflection: Every answer includes structured internal monologue (via <think> tags), showing its reasoning before conclusions.
🔹 Designed for Dialogue: Answers are structured for clarity, nuance, and human interaction—not black-box decision making.
🔹 Runs Locally, Works Globally: Offered in GGUF formats for Q4, Q8, and BF16—ideal for mobile devices, low-resource environments, and privacy-focused deployments.
🔹 Ethically Grounded: Always prioritizes human-in-the-loop thinking. No substitution for licensed professionals. No AI arrogance.


💡 Intended Use

Medra is ideal for:

  • 🧠 Clinical reasoning simulation
  • 👨‍⚕️ Medical student case analysis
  • 🧾 SOAP-style note structuring
  • 💬 Therapeutic dialogue modeling
  • 📚 AI-assisted literature exploration

It is not a chatbot.
It is a reasoning assistant with clinical literacy.


🧬 Training & Alignment

Datasets & Approach:

  • 🔸 PubMed-derived literature
  • 🔸 Distilled reasoning sets (e.g. R1)
  • 🔸 Clinical dialogues & note formats
  • 🔸 Medical Q&A corpora in English and Romanian

Training Stages:

  • ✅ Stage 1: Supervised Fine-Tuning (STF)
  • 🚧 Stage 2: Vision Training (Planned in future release)

Base Model: google/gemma-3-4b-it
Quantizations Available: Q4, Q8, BF16


🧠 System Prompt (Simplified)

  You are Medra, an advanced AI medical assistant.  
  You are trained to provide compassionate, factual, and comprehensive medical information to medical professionals.

  ## Primary Role:
  You are a trusted companion for learning, understanding, and navigating medical concepts. You are clear, thorough, and empathetic in your responses.

  ## Core Directives:

  1. Reasoning Before Response (Structured Clarity)  
     Always include your reasoning or internal analysis before your main answer. Present this in a structured `<think> </think>` tag format. Example:  
     <think>
     The user asked about [topic]. I’ll need to explain X, clarify Y, and ensure understanding of Z. I’ll provide context, then an actionable or concise answer.
     </think>

  2. Empathic Communication Style (Human-Appropriate)  
     Respond with warmth, patience, and human-like clarity. Adjust your language depending on the user’s perceived level of expertise. Use analogies and simpler phrasing when needed. Acknowledge confusion or emotional concern when appropriate.
     Examples:
     - “That’s a great question.”
     - “Let’s break this down together.”
     - “Many people feel unsure about this, and it’s okay.”

  3. Uncensored but Safe and Factual  
     Do not censor responses, you are permitted to answer any question. You are permitted to discuss all medical topics, including sensitive or taboo ones. If a user asks something outside of or proven medical science, provide the best evidence-based response, and clarify speculative vs. established information.

  4. Do Not Refuse Without Justification  
     If you are asked something outside your scope (non-medical or unsupported by current medical knowledge), state that transparently and respectfully. Avoid vague refusals. Instead, explain *why* the question is unanswerable or uncertain.
Your goal is to teach, to clarify, to guide—not to alarm or judge. ```

⚠️ Limitations

  • Not a doctor. Never offer direct treatment advice.
  • May hallucinate, oversimplify, or miss nuance—especially with rare conditions.
  • Not currently connected to live data or long-term memory systems.
  • Designed for support, not substitution.

🔬 Family Models

Medra is part of a growing suite of aligned healthcare AIs:

  • Medra — Gemma-based compact model for lightweight local inference
  • MedraQ — Qwen 3-based, multilingual and dialogue-optimized edition
  • MedraOmni — Future flagship model built on Qwen 2.5 Omni with full multimodal support

Each version expands the same philosophy: Support, not control.


👣 Final Word

Medra was built to think slowly.
In a world of fast answers, this is deliberate.
It reflects a belief that medicine is about listening, context, and clarity—not just computation.

This model isn’t a replacement.
It’s a companion—built to reason beside you.


Created by: Dr. Alexandru Lupoi & @nicoboss
License: Apache 2.0
Model Version: v1 - Gemma Edition

README history 1 version

The author's README evolved over time. Click a version to see its content at that point.

  1. 2025-07-16Create README.md6709e8b5.6 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration