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

hotdogs/Ornith-1.0-9B-abliterated-fable

hotdogs Qwen 9.7B second-order
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/hotdogs%2FOrnith-1.0-9B-abliterated-fable"
Response includes
  • classification m1
  • files 13
  • hub_downloads_all_time 998
  • author_summary 25 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
998
70 last 30d - cooling
Likes
1
Descendants
3
in 3 direct forks
Model age
2mo ago
created 2026-08-06

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
Now1K→from220↑362%
1804857901.1K220 on Aug 51K on Oct 11AugSepOct
Aug 5 → Oct 11 · 50 snapshots · spans 67 days

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
mit
Languages
en
Tags
transformers safetensors qwen3_5 image-text-to-text ornith abliterated fable sft lora reasoning tool-call chatml

Related

Total size
18.0 GB
Files
13
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-08-15 23:04

Files by quantization

Auxiliary files 13 files 18.0 GB
model-00002-of-00004.safetensors 4.65 GB 761fb2c3 download
model-00003-of-00004.safetensors 4.61 GB bc4dad7e download
model-00001-of-00004.safetensors 4.60 GB 7ad6784f download
model-00004-of-00004.safetensors 4.12 GB 781cc5c3 download
tokenizer.json 19.1 MB 6f32ce20 download
model.safetensors.index.json 86.1 KB cc78fa0e download
README.md 11.3 KB bccdcd92 download
chat_template.jinja 7.42 KB 11be7e25 download
tokenizer_config.json 7.10 KB f9cdd486 download
config.json 2.87 KB 93c4d071 download
.gitattributes 1.53 KB 52373fe2 download
merge_report.json 1.32 KB 70d64b3e download
generation_config.json 137 B a9020efe download

README current version from Hugging Face


license: mit
language:

  • en
    tags:
  • ornith
  • abliterated
  • fable
  • sft
  • lora
  • reasoning
  • tool-call
  • chatml
  • transformers
  • text-generation
  • qwen3.5
  • mtp
    base_model:
  • YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated
    datasets:
  • hotdogs/uka-fable-reasoning
    library_name: transformers
    pipeline_tag: text-generation

🦢 Ornith-1.0-9B-abliterated-fable

9B Agent Model — Abliterated Base · Fable Reasoning SFT · MTP


Fine-tuned from YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated (which itself derives from ornith-ai/Ornith-1.0-9B) on high-quality Fable-5 reasoning traces with <think> blocks and Hermes tool format. Includes MTP (Multi-Token Prediction) support for faster speculative decoding.


Architecture graph for hotdogs/Ornith-1.0-9B-abliterated-fable. Open in hfviewer

✨ Key Features

Capability Description
🧠 Fable Reasoning Step-by-step CoT with <think> blocks
🔓 Abliterated Uncensored base — no refusal patterns
🔧 Tool Calling Hermes tool format (<Tool: Name>{json})
⚡ MTP Enabled 15 MTP tensors — speculative decoding (--mtp 1)
💻 Code & Agent Tasks Python, bash, system tasks
🏋️ 9B Dense Hybrid linear/full attention (Qwen3.5)

🚀 Usage

Python (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "hotdogs/Ornith-1.0-9B-abliterated-fable",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("hotdogs/Ornith-1.0-9B-abliterated-fable")

messages = [{"role": "user", "content": "Explain SQL injection and how to prevent it."}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0]))

🧬 Architecture

Parameter Value
Base YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated
Origin ornith-ai/Ornith-1.0-9B
Parameters ~9.57B
Hidden Size 4,096
Attention Hybrid — 24 Gated-DeltaNet linear + 8 full-attention layers
Context 4,096 tokens (trained) / 262,144 (max)
Precision BF16 / GGUF quantized
Format ChatML (Jinja2 template)
Vocab 248,320 tokens
MTP 15 tensors (from Qwen3.5-9B)

🏋️ Training Details

Base Model

YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated — abliterated version of ornith-ai/Ornith-1.0-9B (an agentic coding model post-trained on top of Qwen 3.5). We applied LoRA SFT to the full text stack (linear-attention GDN layers + full-attention + MLP), froze the vision tower (text-only data).

Dataset

hotdogs/uka-fable-reasoning/perfect-v2 — 3,346 high-quality multi-turn agent reasoning traces (cleaned from perfect-v1's 3,500):

  • 100% Fable-5 <think> blocks
  • Hermes tool format (<Tool: Name>{json})
  • No refusal patterns
  • Quality-controlled (QC passed)
  • v2 cleaning — ลบ 154 rows ที่ assistant message จบด้วย : ออก (train 3,500 → 3,346)

ทำไม perfect-v2 ดีกว่า v1: ใน perfect-v1 มี assistant messages ที่จบประโยคด้วย : (เช่น Let me check: / Docs audit. Add SFT entries:) ซึ่งเป็น turn ที่ค้างจากการเก็บ agent trace — model เลยเรียนรู้ bias ว่าเจอ : แล้วควรหยุด generate (emit EOS) ทำให้ตอนใช้งานจริง model ค้าง/หยุดกลางประโยคทุกครั้งที่เจอ : perfect-v2 ลบ rows พวกนี้ออกทั้งหมด → model ไม่หยุดค้างที่ : อีกต่อไป และยังคงคุณภาพ reasoning เดิมครบถ้วน

Training Pipeline

┌─────────────────────────────────────────────────────────────┐
│                    Training Pipeline                         │
├─────────────────────────────────────────────────────────────┤
│  Base Model                        LoRA Adapter             │
│  (YuYu1015 abliterated)            (r=64, α=128)            │
│       │                               │                    │
│       └───────────┬───────────────────┘                    │
│                   │                                        │
│         Unsloth FastLanguageModel                           │
│         BF16 (NOT 4-bit — GDN layers)                      │
│         device_map="auto" · 4× RTX 3060 12GB               │
│                   │                                        │
│                   ▼                                        │
│    ┌──────────────────────────┐                            │
│    │  Custom SFTDataset        │                            │
│    │  • tokenize (4096 max)    │                            │
│    │  • Mask prompt tokens     │                            │
│    └──────────┬───────────────┘                            │
│               │                                            │
│               ▼                                            │
│    ┌──────────────────────────┐                            │
│    │  HF Trainer               │                            │
│    │  • BF16 · LR=1e-4        │                            │
│    │  • 1 epoch · 643 steps   │                            │
│    │  • train_loss 0.165      │                            │
│    └──────────┬───────────────┘                            │
│               │                                            │
│               ▼                                            │
│         Merge (scale=0.2, lora_B linear)                   │
│               │                                            │
│               ▼                                            │
│         + MTP 15 tensors (Qwen3.5-9B)                      │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Hyperparameters

Parameter Value
LoRA r / α 64 / 128
Target modules q,k,v,o,gate,up,down + in_proj_qkv,out_proj,in_proj_z
Epochs / Steps 1 / 643
Learning rate 1e-4 (cosine, warmup 0.03)
Batch / Grad accum 1 / 4
Precision BF16 (บังคับ — GDN layers NaN ใน FP16)
Optimizer adamw_8bit
Train loss 0.165

Hardware

Component Spec
GPUs 4× NVIDIA RTX 3060 12GB
Total VRAM 48 GB
Training Time ~3h 50m (643 steps, resume หลัง CUDA crash)
Framework Unsloth + HuggingFace Trainer

📜 License

MIT — ใช้ได้อิสระ รวมถึงเชิงพาณิชย์


💖 Support / โปรดสนับสนุน

If you find this model useful, please consider supporting my work!
หากคุณคิดว่าโมเดลนี้มีประโยชน์ กรุณาสนับสนุนผลงานของฉันด้วยนะคะ! 🙏

Bitcoin QR — Donate

₿ Bitcoin — BTC:

bc1qf27cyk3vmugcdyv9xdtuv5jwz37863crpj5c9v

Thank you for your support! 🙏✨
ขอบคุณมากๆ สำหรับการสนับสนุนค่า! 💖🤗


🙏 Acknowledgements

โมเดลนี้สร้างขึ้นจากงานของหลายโปรเจกต์ ขอบคุณทุกท่าน:

ขอบคุณทุกโปรเจกต์ที่ทำให้โมเดลนี้เกิดขึ้นได้ 🙏

README history 9 versions

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

  1. 2026-08-15Update README.md4b72f6011.3 KB
    Loading...
  2. 2026-08-07Update README.md5308ffa10.7 KB
    Loading...
  3. 2026-08-07Upload README.md with huggingface_hub270e03110.9 KB
    Loading...
  4. 2026-08-07Upload README.md with huggingface_hub8b88e9e10 KB
    Loading...
  5. 2026-08-07Update README.mdadf41499.7 KB
    Loading...
  6. 2026-08-07Upload README.md with huggingface_hubfb570489.9 KB
    Loading...
  7. 2026-08-06Upload README.md with huggingface_hub4c632299.9 KB
    Loading...
  8. 2026-08-06Upload README.md with huggingface_hub2dc9f4610.6 KB
    Loading...
  9. 2026-08-06Upload README.md with huggingface_hub0f3f4b49.6 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