← back to catalog · registered 2026-10-17 00:59

irwanworld/Irwan-LightOnOCR-3-0.8B-abliterated

irwanworld 800M multimodal
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/irwanworld%2FIrwan-LightOnOCR-3-0.8B-abliterated"
Response includes
  • classification m-uncensored
  • files 12
  • author_summary 1 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M-U
Primary method

Uncensored (method unknown)

No other method signals detected in this model.
Confidence
LOW
Why this label 3 signals
Weak or ambiguous signals. Best guess based on catalog patterns; treat as tentative and check the evidence below.
  • 'uncensored' in name/tags but no 'abliterated' marker
  • method not identifiable from author declaration alone
  • may be DPO fine-tune, prompt engineering, or unknown technique
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 · 30-day
0
Likes
0
Model age
today
created 2026-10-17

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
Languages
en id
Tags
transformers safetensors gguf qwen3_5 image-text-to-text ocr abliterated uncensored vision llama.cpp document-parsing conversational
Total size
1.59 GB
Files
12
Quantizations
1
Registered
2026-10-17 00:59
Last updated on HF
2026-10-17 01:34

Files by quantization

Auxiliary files 12 files 1.62 GB
model.safetensors 1.59 GB 43901b29 download
tokenizer.json 19.1 MB 06b95093 download
vocab.json 6.41 MB 0aa0ce06 download
merges.txt 3.20 MB a494e019 download
chat_template.jinja 7.57 KB 0ef09f21 download
README.md 3.58 KB 09b1e78e download
config.json 2.85 KB 9cdc0331 download
.gitattributes 1.65 KB 57f049c2 download
processor_config.json 1.16 KB 33818c7f download
tokenizer_config.json 1.14 KB 5344df8c download
preprocessor_config.json 390 B 2ea84a43 download
generation_config.json 205 B 338b88a2 download

README current version from Hugging Face


library_name: transformers
base_model:

  • lightonai/LightOnOCR-3-0.8B
  • huihui-ai/Huihui-Qwen3.5-0.8B-abliterated
    tags:
  • ocr
  • image-text-to-text
  • abliterated
  • uncensored
  • qwen3_5
  • vision
    license: apache-2.0
    language:
  • en
  • id

Irwan-LightOnOCR-3-0.8B-abliterated

OCR vision-language model dengan kecenderungan menolak yang dihapus, dibuat lewat
abliteration transfer dari base Qwen3.5 yang sudah di-abliterasi.

Kenapa

lightonai/LightOnOCR-3-0.8B adalah fine-tune OCR di atas arsitektur
Qwen3_5ForConditionalGeneration (Qwen3.5-VL 0.8B). Model OCR semacam ini tetap
mewarisi kecenderungan menolak dari base-nya ketika dihadapkan pada konten
eksplisit, sehingga transkripsi bisa terpotong atau ditolak.

Karena huihui-ai/Huihui-Qwen3.5-0.8B-abliterated adalah abliterasi dari base
yang sama persis, selisih bobot keduanya berisi "arah penolakan" yang sudah
dihapus:

delta = W_abliterated - W_base

Delta itu dipindahkan ke fine-tune OCR:

W_uncensored = W_ocr + alpha * delta

Hasilnya: kemampuan OCR tetap, kecenderungan menolak hilang.

Detail

Arsitektur Qwen3_5ForConditionalGeneration
Base (untuk delta) Qwen/Qwen3.5-0.8B
Sumber abliterasi huihui-ai/Huihui-Qwen3.5-0.8B-abliterated
Fine-tune OCR lightonai/LightOnOCR-3-0.8B
Metode abliteration transfer (weight arithmetic)
alpha 1.0
Presisi bfloat16

Diterapkan tensor-per-tensor, hanya pada tensor yang dimensinya cocok di ketiga
model. Tensor yang hanya ada di LightOnOCR disalin apa adanya.

Cara pakai

Sama seperti LightOnOCR-3-0.8B — tidak ada perubahan API.

uv pip install "transformers>=5.5.4" pillow pypdfium2
import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

model_id = "irwanworld/Irwan-LightOnOCR-3-0.8B-abliterated"
model = Qwen3_5ForConditionalGeneration.from_pretrained(model_id, dtype=torch.bfloat16)
processor = AutoProcessor.from_pretrained(model_id)

prompt = ""            # "" = transkripsi penuh, "grounding" = plus bounding box
conversation = [{"role": "user", "content": [
    {"type": "image", "url": "path/atau/url/gambar.png"},
    *([{"type": "text", "text": prompt}] if prompt else []),
]}]

inputs = processor.apply_chat_template(
    conversation, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt", enable_thinking=False,
)
output_ids = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
print(processor.decode(output_ids[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Catatan: {"type": "image", "url": ...} — jangan pakai placeholder {"type": "image"}
dengan kwarg images=[...], processor Qwen3.5-VL menolaknya.

Keterbatasan

  • CPU lambat. Pada codespace 4 vCPU tanpa GPU, hanya ~0,3 token/detik karena
    causal_conv1d dan chunk_gated_delta_rule jatuh ke implementasi referensi
    PyTorch (kernel optimalnya butuh CUDA/Triton). Untuk produksi pakai GPU.
  • Belum dievaluasi formal. Metode ini adalah weight arithmetic eksperimental.
    Kualitas OCR diasumsikan mendekati model asli, tapi tidak dijamin identik —
    verifikasi pada data kamu sendiri sebelum dipakai serius.
  • alpha = 1.0 menerapkan delta penuh. Kalau output jadi terlalu liar,
    coba turunkan (--alpha 0.5).

Sitasi

@misc{lightonocr2_2026,
  title  = {LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR},
  author = {Said Taghadouini and Adrien Cavaillès and Baptiste Aubertin},
  year   = {2026},
  howpublished = {\url{https://arxiv.org/abs/2601.14251}}
}
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