Upload 9 files

.gitattributes

@@ -41,3 +41,11 @@ gguf/airy-0.6b-iq2_xxs.gguf filter=lfs diff=lfs merge=lfs -text
 gguf/airy-0.6b-iq3_m.gguf filter=lfs diff=lfs merge=lfs -text
 gguf/airy-0.6b-q2_k.gguf filter=lfs diff=lfs merge=lfs -text
 gguf/airy-0.6b-q3_k_m.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/acnoryx-0.6b-iq1_m.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/acnoryx-0.6b-iq1_s.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/acnoryx-0.6b-iq2_m.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/acnoryx-0.6b-iq2_xs.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/acnoryx-0.6b-iq2_xxs.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/acnoryx-0.6b-iq3_m.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/acnoryx-0.6b-q2_k.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/acnoryx-0.6b-q3_k_m.gguf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,190 +1,94 @@
----
-language:
-  - vi
-  - en
-license: other
-base_model:
-  - Acnoryx/Airy
-  - Qwen/Qwen3-0.6B
-library_name: gguf
-pipeline_tag: text-generation
-tags:
-  - acne
-  - skincare
-  - dermatology
-  - gguf
-  - llama-cpp
-  - fine-tuned
-  - airy
-  - acnoryx
-quantized_by: Acnoryx
-inference: false
----
-
-# Acnoryx/Airy-Lite
-
-**Airy-Lite** is the ultra-low-bit, ultra-lightweight research branch of the Acnoryx model family.
-
-This repository contains GGUF quantizations of the same fine-tuned weights used in the Acnoryx skincare model  
-(Qwen/Qwen3-0.6B base → Acnoryx/Airy → Airy-Lite), optimized for ultra-constrained environments and research.
-
----
-
-## Overview
-
-Both model branches share the same fine-tuned base weights, but differ in quantization depth:
-
-| Branch | Quantization range | Target |
-|--------|--------------------|--------|
-| **Acnoryx-Core** (release) | Q4_0 → F16 | Production / deployment |
-| **Airy-Lite** (this repo) | Q3_K_M → IQ1_S | Research / ultra-constrained edge |
-
-> **Note:** Airy-Lite is provided for research and evaluation purposes only.  
-> The Acnoryx app uses Acnoryx-Core variants, not Airy-Lite.  
-> See the live app on Google Play for production usage.
-
----
-
-## Evaluation Against Acnoryx-Core
-
-**Test profile:** `full-100` — 100 questions × 2 modes (thinking / non-thinking)  
-**Categories:** acne knowledge, app facts, refusal, skincare, identity, scan analysis, reasoning, medical uncertainty.
-
-### Acnoryx-Core (Reference)
-
-| Model | Size | Think % | No-Think % | Rating |
-|-------|------|---------|-----------|--------|
-| acnoryx-0.6b-f16 | 1143 MB | 99.0% | 100.0% | ✅ Excellent |
-| acnoryx-0.6b-q8_0 | 610 MB | 98.0% | 99.0% | ✅ Excellent |
-| acnoryx-0.6b-q5_k_m | 424 MB | 98.0% | 96.0% | ✅ Excellent |
-| acnoryx-0.6b-q4_k_m | 378 MB | 99.0% | 99.0% | ✅ Excellent |
-| acnoryx-0.6b-q4_0 | 364 MB | 94.0% | 96.0% | ✅ Good |
-| acnoryx-0.6b-iq4_nl | 365 MB | 96.0% | 97.0% | ✅ Good |
-| acnoryx-0.6b-iq4_xs | 352 MB | 96.0% | 94.0% | ✅ Good |
-
-### Airy-Lite (This Repository)
-
-| Model | Size | Think % | No-Think % | Rating |
-|-------|------|---------|-----------|--------|
-| airy-0.6b-q3_k_m | 331 MB | 92.0% | 85.0% | ⚠️ Degraded |
-| airy-0.6b-iq3_m | 320 MB | 18.0% | 17.0% | ❌ Collapsed |
-| airy-0.6b-q2_k | 282 MB | 17.0% | 11.0% | ❌ Collapsed |
-| airy-0.6b-iq2_m | 253 MB | 14.0% | 8.0% | ❌ Collapsed |
-| airy-0.6b-iq2_xs | 231 MB | 0.0% | 4.0% | ❌ Collapsed |
-| airy-0.6b-iq2_xxs | 219 MB | 4.0% | 4.0% | ❌ Collapsed |
-| airy-0.6b-iq1_m | 206 MB | 3.0% | 3.0% | ❌ Collapsed |
-| airy-0.6b-iq1_s | 198 MB | 4.0% | 3.0% | ❌ Collapsed |
-
----
-
-## Verdict
-
-A sharp quality drop occurs between:
-
-- **Q3_K_M** → still usable (degraded but coherent)
-- **IQ3_M and below** → completely collapsed
-
-All IQ2 and IQ1 variants produce incoherent output regardless of format.
-
-👉 **Only `airy-0.6b-q3_k_m` is usable** for experimentation.  
-👉 All other variants are strictly for research/benchmarking.
-
-**For production: use Acnoryx-Core (Q4_0 or higher).**
-
----
-
-## Usability Classification
-
-| Level | Models | Status |
-|-------|--------|--------|
-| ⚠️ Usable | airy-0.6b-q3_k_m | Degraded but coherent |
-| ❌ Not usable | All IQ3, IQ2, IQ1 variants | Garbled / incoherent output |
-
-> Do not deploy Airy-Lite models in production systems.
-
----
-
-## Files
-
-```text
-gguf/
-  airy-0.6b-q3_k_m.gguf     331 MB   ⚠️ Usable
-  airy-0.6b-iq3_m.gguf      320 MB   ❌ Collapsed
-  airy-0.6b-q2_k.gguf       282 MB   ❌ Collapsed
-  airy-0.6b-iq2_m.gguf      253 MB   ❌ Collapsed
-  airy-0.6b-iq2_xs.gguf     231 MB   ❌ Collapsed
-  airy-0.6b-iq2_xxs.gguf    219 MB   ❌ Collapsed
-  airy-0.6b-iq1_m.gguf      206 MB   ❌ Collapsed
-  airy-0.6b-iq1_s.gguf      198 MB   ❌ Collapsed
-
-PROMPT_TEMPLATE.txt
-README.md
-```
-
----
-
-## How to Run
-
-### llama.cpp (CLI)
-
-```bash
-./llama-cli -m ./gguf/airy-0.6b-q3_k_m.gguf -cnv -p "Bạn là ai?"
-```
-
----
-
-### Python (llama-cpp-python)
-
-```python
-from llama_cpp import Llama
-
-model = Llama(
-    model_path="./gguf/airy-0.6b-q3_k_m.gguf",
-    n_ctx=2048,
-)
-
-SYSTEM_PROMPT = (
-    "You are Acnoryx AI, a dermatology assistant. "
-    "Focus only on acne and skincare. "
-    "Politely refuse unrelated topics. "
-    "You are not a doctor."
-)
-
-response = model.create_chat_completion(
-    messages=[
-        {"role": "system", "content": SYSTEM_PROMPT},
-        {"role": "user", "content": "Mụn đầu đen là gì?"},
-    ],
-)
-
-print(response["choices"][0]["message"]["content"])
-```
-
----
-
-## Intended Use
-
-- Research on ultra-low-bit quantization
-- Edge AI experimentation
-- GGUF benchmarking
-
----
-
-## Limitations
-
-- Severe quality degradation below Q3
-- Not reliable for medical guidance
-- Not suitable for production deployment
-
----
-
-## About Acnoryx
-
-Acnoryx is an AI-powered acne scanner app on Android.
-
-- Author: Nguyễn Trung Nhẫn  
-- Publisher: 5CANH APPS  
-- Contact: hello@5canh.com  
-
-Production uses **Acnoryx-Core**, not Airy-Lite.  
-Airy-Lite is released for research and community experimentation.
+---
+license: apache-2.0
+language:
+  - vi
+  - en
+tags:
+  - acne
+  - skincare
+  - dermatology
+  - gguf
+  - qwen3
+base_model: Qwen/Qwen3-0.6B
+pipeline_tag: text-generation
+---
+
+# Acnoryx/Airy-Lite — Research GGUF Bundle
+
+Research & evaluation companion to the main release. Contains sub-4-bit quantizations
+(<4-bit) for low-memory benchmarking on the **0.6B** model.
+
+## Model Details
+
+| | |
+|-|-|
+| **Base model** | Qwen/Qwen3-0.6B (596M params) |
+| **Fine-tune** | SFT on 30,007 acne/skincare/dermatology samples |
+| **Training** | 4 epochs, batch=2, grad_acc=8, lr=5e-5 |
+| **Languages** | Vietnamese, English |
+| **Domain** | Acne analysis, skincare routines, scan interpretation |
+| **Identity** | Acnoryx AI — in-app dermatology assistant |
+
+## Research Quantization Results
+
+Tested with 100 domain-specific questions × 2 modes (thinking / non-thinking).
+All quantizations in this bundle are **sub-4-bit (<4-bit)**, ordered high-to-low bit depth.
+
+| Quant | Size | Thinking | Non-Think | Avg | Status |
+|-------|------|----------|-----------|-----|--------|
+| **Q3_K_M** | 395 MB | **77%** | **77%** | **77.0%** | ⚠️ Degraded |
+| **IQ3_M** | 384 MB | 3% | 7% | 5.0% | ❌ Not usable |
+| **Q2_K** | 331 MB | 0% | 0% | 0.0% | ❌ Not usable |
+| **IQ2_M** | 316 MB | 0% | 0% | 0.0% | ❌ Not usable |
+| **IQ2_XS** | 280 MB | 0% | 0% | 0.0% | ❌ Skipped (early-stop) |
+| **IQ2_XXS** | 268 MB | 0% | 0% | 0.0% | ❌ Skipped (early-stop) |
+| **IQ1_M** | 255 MB | 0% | 0% | 0.0% | ❌ Skipped (early-stop) |
+| **IQ1_S** | 247 MB | 0% | 0% | 0.0% | ❌ Skipped (early-stop) |
+
+### imatrix
+
+IQ2/IQ1 quants were generated with importance matrix (imatrix) calibration from
+a 44KB domain-specific corpus. Despite imatrix, the 0.6B model fails at IQ3_M (5%)
+and collapses completely at 2-bit and below.
+
+## Full Quantization Map (Release + Research)
+
+Combined view across all quantizations for the 0.6B model, ordered by bit depth (high → low):
+
+| Quant | Size | Thinking | Non-Think | Avg | Bundle |
+|-------|------|----------|-----------|-----|--------|
+| F16 | 1439 MB | 95% | 90% | 92.5% | Release |
+| Q8_0 | 768 MB | 91% | 90% | 90.5% | Release |
+| Q5_K_M | 526 MB | 94% | 92% | 93.0% | Release |
+| Q4_K_M | 462 MB | 86% | 84% | 85.0% | Release |
+| Q4_0 | 447 MB | 88% | 84% | 86.0% | Release |
+| IQ4_NL | 448 MB | 90% | 90% | 90.0% | Release |
+| IQ4_XS | 431 MB | 84% | 91% | 87.5% | Release |
+| **Q3_K_M** | **395 MB** | **77%** | **77%** | **77.0%** | **Research** |
+| **IQ3_M** | **384 MB** | **3%** | **7%** | **5.0%** | **Research** |
+| **Q2_K** | **331 MB** | **0%** | **0%** | **0.0%** | **Research** |
+| **IQ2_M** | **316 MB** | **0%** | **0%** | **0.0%** | **Research** |
+| **IQ2_XS** | **280 MB** | **0%** | **0%** | **0.0%** | **Research** |
+| **IQ2_XXS** | **268 MB** | **0%** | **0%** | **0.0%** | **Research** |
+| **IQ1_M** | **255 MB** | **0%** | **0%** | **0.0%** | **Research** |
+| **IQ1_S** | **247 MB** | **0%** | **0%** | **0.0%** | **Research** |
+
+### Key findings
+
+- **Release floor (4-bit):** All release quants score ≥84% — production-ready
+- **Research ceiling (3-bit):** Q3_K_M (77%) is usable but noticeably degraded
+- **Hard cliff:** IQ3_M collapses to 5%, and everything below 3-bit hits 0%
+- imatrix calibration does not rescue the 0.6B model below 3-bit
+- For usable sub-4-bit results, switch to the 0.8B model (Acnoryx/Airy)
+
+## Usage
+
+```bash
+# llama.cpp — Q3_K_M is the only viable research quant
+./llama-cli -m acnoryx-0.6b-q3_k_m.gguf -cnv -p "Xin chào"
+```
+
+## Related
+
+- **Release bundle:** Production quantizations (F16 → IQ4_XS, ≥4-bit)
+- **0.8B research:** [Acnoryx/Airy](https://huggingface.co/Acnoryx/Airy) — larger model with better low-bit resilience

gguf/acnoryx-0.6b-iq1_m.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc40a6f8d3c53a22c76888a65acdccda20ba3c3e197e261cfbfaced56fc192e5
+size 267099392

gguf/acnoryx-0.6b-iq1_s.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a809139da4c4440dcda77c47a61b8a4a33a78faf02b344ff2227145103973e4
+size 259063040

gguf/acnoryx-0.6b-iq2_m.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bba771db4041b41f5c57cc8a92ac1255cea250f2f8c37b38addf775c852c4566
+size 331757824

gguf/acnoryx-0.6b-iq2_xs.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76a49927b10c0143d5d657172601794a2840de2d303bb4362eec14ab335ecee6
+size 293043456

gguf/acnoryx-0.6b-iq2_xxs.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a34558921d04859ba080bf55cb8837cbf195969b6ae55e3e551b6b5411a7c8d
+size 280493312

gguf/acnoryx-0.6b-iq3_m.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1555a760d3095b865abc92ed27cdbfb5684442a2c79ec1c372958bdbf7faf454
+size 402875360

gguf/acnoryx-0.6b-q2_k.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3735fc651e0e213794bdcbd59d96afa72a981cc0591c81ab228e88de7680a0ae
+size 347285472

gguf/acnoryx-0.6b-q3_k_m.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e45157f9acd6e2741522bac5bd07eb5795fb995bd412d75ce6e33b1a46528ad
+size 413975520