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

Krypto-Whitehat/qwen3.8-9b-uncensored-cyber-xrpl-v4

Krypto-Whitehat Qwen 9B GGUF second-order 262K 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/Krypto-Whitehat%2Fqwen3.8-9b-uncensored-cyber-xrpl-v4"
Response includes
  • classification m8
  • files 7
  • hub_downloads_all_time 642
  • author_summary 3 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M8
Primary method

Repackaging (quantization)

Applied on top of direct removal inherited from the base model.
Confidence
MEDIUM
Inherited from base model
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=1
  • assume M1 (base ablation) + M8 (GGUF quant) - default when producer unknown
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · lifetime
642
214 last 30d - stable
Likes
1
Model age
7w ago
created 2026-08-22
Downloads over time
Now715→from0↑0%
02625247870 on Aug 19715 on Oct 12AugSepOct
Aug 19 → Oct 12 · 49 snapshots · spans 54 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.

Metadata

License
apache-2.0
Quantizations
Q6_K Q8_0
Tags
gguf xrpl cybersecurity anti-hallucination qlora qwen3.5 q6_k q8_0 abliterated text-generation base_model:MegaPanchamZ/Qwen3.8-9B-abliterated-25 base_model:quantized:MegaPanchamZ/Qwen3.8-9B-abliterated-25

Related

Total size
15.7 GB
Files
7
Quantizations
3
Registered
2026-08-22 13:56
Last updated on HF
2026-08-22 23:10

Files by quantization

Q8_0 1 file 8.87 GB
model.Q8_0.gguf 8.87 GB 1a7bb17d download
Q6_K 1 file 6.85 GB
model.Q6_K.gguf 6.85 GB 03e54784 download
Auxiliary files 5 files 4.92 MB
imatrix.dat 4.91 MB 9787ca6f download
README.md 4.44 KB 557a7765 download
hf_upload_manifest.json 2.04 KB 52b88275 download
.gitattributes 1.63 KB ddc57bb0 download
Modelfile 273 B fc1c0525 download

README current version from Hugging Face


license: apache-2.0
base_model: MegaPanchamZ/Qwen3.8-9B-abliterated-25
tags:

  • xrpl
  • cybersecurity
  • anti-hallucination
  • qlora
  • abliterated
  • qwen3.5
  • gguf
  • q6_k
  • q8_0
    pipeline_tag: text-generation

Krypto-Whitehat/qwen3.8-9b-uncensored-cyber-xrpl-v4

QLoRA-SFT von MegaPanchamZ/Qwen3.8-9B-abliterated-25 (Heretic-abliterated, BF16 Safetensors) auf einen anti-halluzinierenden XRPL/Cyber-Korpus. Trainiert auf RTX 5090 Laptop (24 GB GDDR7), 2 Epochen, lora_r=16, eval_strategy=epoch.

Empfohlener Quant

Datei Größe Rolle
model.Q6_K.gguf 6.9 GB HF-Default — beste Balance für 24 GB-GPUs
model.Q8_0.gguf 8.9 GB Qualitäts-Referenz, nahe F16
model.F16.gguf 17 GB Referenz / Reproduzierbarkeit, zu groß als Default

Q4_K_M ist absichtlich NICHT enthalten — für strukturierte Cyber/VERDICT/file:line-Ausgabe ist die Qualitätseinbuße spürbar.

Basis

  • Base: MegaPanchamZ/Qwen3.8-9B-abliterated-25 (Heretic Decensoring, BF16 Safetensors)
  • Training: QLoRA 4-bit NF4 + LoRA r=16, alpha=32, alle q/k/v/o + MLP auf ausgewählten Schichten (unsloth-Filter)
  • Trainingsdaten: mix_sft_train.jsonl (184 Rows: 112 XRPL + 72 externe Cyber-Quelle)
  • Eval: mix_sft_eval.jsonl (33 Rows), train_loss=1.691, eval_loss=1.428 nach 2 Epochen
  • Holdout (nur für imatrix, nicht für SFT): corpus/imatrix_holdout.jsonl (80 Rows, user-Prompts)

Quantisierung

  • Toolchain: llama.cpp CUDA-Build (sm_120 = RTX 5090 Blackwell)
  • imatrix: berechnet aus corpus/imatrix_holdout.jsonl (98 Prompts × 2 dupliziert = 196 Zeilen, ~3700 Tokens, c=1024)
  • Imatrix-SHA256[:16]: siehe adapter/gguf/imatrix.dat
  • Q8_0: 8,50 BPW, 8.9 GB, sha256[:16]=1a7bb17d98f17155
  • Q6_K: 6,57 BPW, 6.9 GB, sha256[:16]=03e547841be07d73
  • F16: 16,00 BPW, 17 GB, sha256[:16]=15d5b92ab7ddd9aa

Alle drei GGUFs: 427 Tensors, 33 attention_layer (qwen3.5 hybrid: Mamba-SSM mit full_attention_interval=4).

Anti-Halluzinations-Korpus

  • 9 Familien: correctness, cross-feature, develop-only, dos, hygiene, logic-flaws, safety-n11, sentinel-zero, validation-bypass
  • VERDICT-Pflicht: Jede Antwort muss VERDICT: EXPLOITABLE | UNPROVEN | SAFE | NEEDS_REVIEW enthalten
  • File:line-Rate: 100 % der Trainings-Rows mit Zitaten
  • Gold-Eval: G1–G8 nur in Eval, nie in Train
  • Qualitäts-Gate: python quality_gate.py → QUALITY GATE PASSED

Hinweise zur Nutzung

  • LM Studio: Q6_K.gguf laden, GPU-Offload hoch, Context 4k–8k reicht für den Korpus-Stil
  • Chat-Template: Qwen3.8 / ChatML (<|im_start|>...<|im_end|>) — bereits in der GGUF eingebettet
  • Wenn keine Live-PoC verfügbar: das Modell antwortet mit VERDICT: UNPROVEN statt zu halluzinieren
  • Bei Delegation an externes Tooling: Token xrpl-drill wird im Output erwartet

Bekannte Einschränkungen

  • llama.cpp-qwen3.5-CUDA-Pfad ist experimentell; Inferenz via transformers/unsloth/ollama/vllm ist der robustere Weg
  • Smoke-Test auf GGUF via llama-cli wurde nicht im vollen Forward-Pass durchlaufen — strukturelle Validierung (sha256, Tensor-Count, Metadaten) ist sauber, Inferenz-Verifikation via LM Studio/vllm steht aus
  • Eval-Set ist klein (33 Rows); für Produktiv-Einsatz ist eine breitere Anti-Halluzination-Evaluation (z. B. Gold-Set mit 200+ Rows) empfohlen

README history 2 versions

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

  1. 2026-08-22Fix card: VERDICT schema, file:line nuance, no F16/F4 claim, Modelfile FROM path2b0c7634.4 KB
    Loading...
  2. 2026-08-22Upload README.md with huggingface_hubf2d44183.3 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