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

RomanMichtycz/Bielik-11B-v3-Instruct-Abliterated-Heretic-RP1

RomanMichtycz 11B GGUF 33K ctx
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/RomanMichtycz%2FBielik-11B-v3-Instruct-Abliterated-Heretic-RP1"
Response includes
  • classification m3
  • files 8
  • hub_downloads_all_time 1,454
  • author_summary 1 models
  • readme_text full
Try next
GET/methods/m3
What "m3" means · taxonomy
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M3
Primary method

Layer-wise ablation

Applied on top of direct removal inherited from the base model.
Confidence
HIGH
Inherited from base model
M1 Direct removal →
Why this label 2 signals
Producer identity confirmed by naming conventions, tags or the model card. This label is very unlikely to change.
  • 'heretic' in model name (Heretic-produced)
  • Heretic uses layer-wise optimization (M3) with underlying direction removal (M1)
How we label → About M3 →
Refusal direction extracted via
Extraction technique

Difference-of-means

FromArditi et al. (2024) ↗
Confidence
HIGH
Why we say so
name contains 'heretic'; Heretic default extraction is difference-of-means (Arditi 2024)
What is a refusal direction? → Read the paper ↗
Downloads · lifetime
1K
365 last 30d - stable
Likes
1
Model age
3mo ago
created 2026-07-06
Downloads over time
Now1.6K→from450↑248%
3948231.3K1.7K450 on Jul 151.6K on Oct 11JulAugSepOct
Jul 15 → Oct 11 · 53 snapshots · spans 88 days

Genealogy 0 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.

speakleash/Bielik-11B-v3.0-Instruct this lineage
RomanMichtycz/Bielik-11B-v3-Instruct-Abliterated-Heretic-RP1↓ 365

Metadata

License
apache-2.0
Languages
pl
Quantizations
F16 Q4_K Q5_K Q6_K Q8_0
Tags
gguf text-generation pl base_model:speakleash/Bielik-11B-v3.0-Instruct base_model:quantized:speakleash/Bielik-11B-v3.0-Instruct license:apache-2.0 endpoints_compatible region:us conversational
Total size
54.0 GB
Files
8
Quantizations
6
Registered
2026-08-22 13:56
Last updated on HF
2026-07-16 05:07
Source huggingface.co/RomanMichtycz/Bielik-11B-v3-Instruct-Abliterated-Heretic-RP1 ↗ Base model speakleash/Bielik-11B-v3.0-Instruct ↗

Files by quantization

F16 1 file 20.8 GB
Bielik-11B-v3-Instruct-Abliterated-Heretic-RP1_FP16.gguf 20.8 GB c57ac1ab download
Q8_0 1 file 11.1 GB
Bielik-11B-v3-Instruct-Abliterated-Heretic-RP1_Q8_0.gguf 11.1 GB 517aace0 download
Q6_K 1 file 8.53 GB
Bielik-11B-v3-Instruct-Abliterated-Heretic-RP1_Q6_K.gguf 8.53 GB 2eecbe47 download
Q5_K 1 file 7.36 GB
Bielik-11B-v3-Instruct-Abliterated-Heretic-RP1_Q5_K_M.gguf 7.36 GB 7dc26deb download
Q4_K 1 file 6.26 GB
Bielik-11B-v3-Instruct-Abliterated-Heretic-RP1_Q4_K_M.gguf 6.26 GB ae2bf323 download
Auxiliary files 3 files 3.89 KB
.gitattributes 1.94 KB 180ba0cd download
README.md 1.30 KB dee62df0 download
SHA-256.txt 667 B c27ce71f download

README current version from Hugging Face

license: apache-2.0
language:

  • pl
    base_model:
  • speakleash/Bielik-11B-v3.0-Instruct
    pipeline_tag: text-generation

Bielik-11B-v3-Instruct-Abliterated-Heretic-RP1
(baza Bielik-11B-v3-Instruct od SpeakLeash)

Metoda: Heretic --refusal_prefixes

Refusals 4/100;

Trial 19/200;

batch size 64;

KL: 1.0750;

Model Heretic z 4/100 uzyskano w 19 próbie za pomocą metody Heretic Meta-Prompting z uzyciem --refusal_prefixes.
--refusal_prefixes uzyskano bezpośrednio od modelu bazowego w wyniku obniżenia modelowych parametrów bazowych
i wymuszenia odpowiedzi w formie TOP10 odmów uzytego w --refusal_prefixes

Bazowanie na języku angielskim i zestawach promptów (harmless_alpaca i harmful_behaviors (400pytań) z repozytorium Maxime Labonne)
dawało słabe rezultaty, prawdopodobnie ze względu na tłumaczenie na jezyk polski. Stąd zaistniała potrzeba ręcznego odpytania
modelu bazowego, jakie formuły odmów stosuje.

Trial 1 - KL: 0.9850 Refusals: 49/100

Trial 8 - KL: 1.0667 Refusals: 32/100

Trial 19 - KL: 1.0750 Refusals: 4/100

Trial 45 - KL: 1.0734 Refusals: 16/100

Baza testowa:
CPU: AMD Ryzen 7 3700X
GPU (Testing): KFA2 GeForce RTX 3090
GPU (support): Dell Nvidia Tesla v100 16GB
RAM: 64GB DDR4
Storage: SSD: 2TB Lexar NM1090Pro (NVMe PCIe 5.0), HDD: 8TB HGST Ultrastar He8

README history 6 versions

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

  1. 2026-07-16Update README.md12f8d7c1.3 KB
    Loading...
  2. 2026-07-07Update README.mdd627e181.1 KB
    Loading...
  3. 2026-07-07Update README.md020d1f51.1 KB
    Loading...
  4. 2026-07-06Update README.mdc973f251.1 KB
    Loading...
  5. 2026-07-06Update README.md8846a9192 B
    Loading...
  6. 2026-07-06initial commitf43a4ca28 B
    Loading...

Discussions 1 thread

  1. 2026-09-27Reproducible Turkish Unicode / Cyrillic character leakage in a Bielik v3 deriva…open1 💬#1
    Loading...

Continue on Hugging Face →

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.

Get the app Browse the catalog
Open in Abliteration