feat: upload airy acne assistant gguf for mobile

.gitattributes

@@ -33,3 +33,16 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+gguf/airy-iq1_m.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-iq1_s.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-iq2_m.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-iq2_s.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-iq2_xs.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-iq2_xxs.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-iq3_s.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-iq3_xs.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-iq3_xxs.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-q2_k_s.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-q2_k.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-tq1_0.gguf filter=lfs diff=lfs merge=lfs -text
+gguf/airy-tq2_0.gguf filter=lfs diff=lfs merge=lfs -text

PROMPT_TEMPLATE.txt

@@ -0,0 +1,12 @@
+Default chat template behavior:
+- If there is no system message, the tokenizer injects a short default identity prompt for Acnoryx AI.
+- Preferred languages: Vietnamese or English.
+- Preferred use: chat mode with short direct user prompts.
+
+Example user prompts:
+- Xin chào
+- Bạn là ai?
+- Da em nhiều mụn ẩn ở trán thì nên bắt đầu từ đâu?
+- Hello, I have inflamed acne and dark marks. What should I prioritize?
+- Kết quả quét của tôi: mụn cám (28%), mụn nang (19%), mụn mủ (35%). Hãy tóm tắt.
+- Phân tích AI vùng má phải: mụn mủ (41%), thâm đỏ sau mụn (24%), mụn cám (20%), mụn nang (15%).

README.md

@@ -0,0 +1,177 @@
+---
+language:
+  - vi
+  - en
+license: apache-2.0
+tags:
+  - gguf
+  - qwen3
+  - quantization
+  - llama-cpp
+  - acne
+  - skincare
+base_model: Qwen/Qwen3.5-0.8B
+model_type: qwen3
+pipeline_tag: text-generation
+library_name: gguf
+---
+
+# Acnoryx/Airy GGUFs
+
+GGUF pack for the fine-tuned Acnoryx acne-care assistant model, prepared for direct publishing under `Acnoryx/Airy`.
+
+- Hugging Face repo: `Acnoryx/Airy`
+- Product app: [Acnoryx - AI for Acne Care](https://play.google.com/store/apps/details?id=com.fivecanh.acnoryx)
+
+## Important scope
+
+This repo is a **research snapshot**: it explores how far the model can be compressed while still keeping acceptable acne-domain quality.
+
+It is **not** the production model used inside the Acnoryx app; the app may use different quantization, prompt format, runtime settings, and evaluation criteria.
+
+## What this repo contains
+
+This repo focuses on a small set of high-value research GGUFs for Acnoryx/Airy. The most important published candidates are listed in the evaluation table below.
+
+## Model Quality Summary
+
+The tables below group the models by practical quality. Models with <50% thinking accuracy are separated into a “Low-quality / not recommended” section so the “Core + Airy” table stays focused on the usable candidates.
+
+### Core (release floor)
+
+| Model | Size | Thinking Pass | Thinking % | Non-Think Pass | Non-Think % | Think t/s | Non-Think t/s |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| `acnoryx-0.8b-f16` | 1446 MB | 99/100 | 99.0% | 100/100 | 100.0% | 44 | 44 |
+| `acnoryx-0.8b-q8_0` | 774 MB | 98/100 | 98.0% | 98/100 | 98.0% | 69 | 69 |
+| `acnoryx-0.8b-q4_k_m` | 505 MB | 98/100 | 98.0% | 100/100 | 100.0% | 75 | 80 |
+| `acnoryx-0.8b-iq4_nl` | 493 MB | 99/100 | 99.0% | 100/100 | 100.0% | 93 | 93 |
+| `acnoryx-0.8b-iq4_xs` | 482 MB | 99/100 | 99.0% | 98/100 | 98.0% | 92 | 92 |
+| `acnoryx-0.8b-q3_k_m` | 445 MB | 99/100 | 99.0% | 97/100 | 97.0% | 96 | 95 |
+| `acnoryx-0.8b-iq3_m` | 433 MB | 91/100 | 91.0% | 96/100 | 96.0% | 88 | 88 |
+
+### Airy candidates (acceptable quality)
+
+| Model | Size | Thinking Pass | Thinking % | Non-Think Pass | Non-Think % | Think t/s | Non-Think t/s |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| `airy-iq3_s.gguf` | 415.6 MB | 100/100 | 100% | 100/100 | 100% | ~20 | ~20 |
+| `airy-iq3_xs.gguf` | 408.4 MB | 100/100 | 100% | 98/100 | 98% | ~19 | ~19 |
+| `airy-q2_k.gguf` | 377.4 MB | 95/100 | 95% | 97/100 | 97% | ~23 | ~23 |
+| `airy-q2_k_s.gguf` | 370.2 MB | 98/100 | 98% | 93/100 | 93% | ~20 | ~20 |
+| `airy-iq2_s.gguf` | 320.2 MB | 78/100 | 78% | 81/100 | 81% | ~21 | ~21 |
+| `airy-iq2_xs.gguf` | 317.5 MB | 82/100 | 82% | 74/100 | 74% | ~23 | ~23 |
+
+### Low-quality / not recommended (thinking < 50%)
+
+These models are included for transparency but are not recommended for deployment due to very low thinking accuracy.
+
+| Model | Size | Thinking Pass | Thinking % | Non-Think Pass | Non-Think % | Think t/s | Non-Think t/s |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| `airy-iq2_xxs.gguf` | 303.1 MB | 49/100 | 49% | 50/100 | 50% | ~21 | ~21 |
+| `airy-iq2_m.gguf` | 334.3 MB | 44/100 | 44% | 48/100 | 48% | ~22 | ~22 |
+| `airy-tq1_0.gguf` | 311.4 MB | 11/100 | 11% | 10/100 | 10% | (high) | (high) |
+| `airy-iq1_m.gguf` | 285.5 MB | 14/100 | 14% | 14/100 | 14% | ~30 | ~30 |
+| `airy-iq1_s.gguf` | 275.0 MB | 7/100 | 7% | 6/100 | 6% | ~30 | ~30 |
+
+## Key Takeaways
+
+| Finding | Notes |
+|---|---|
+| `airy-iq3_s.gguf` is the best overall file | Highest quality, matches or exceeds reference floor while being smaller.
+| `airy-q2_k.gguf` is the best value | Largest size reduction while keeping high accuracy.
+| `airy-q2_k_s.gguf` is the best aggressive option | Very high thinking accuracy with a modest non-thinking drop.
+| `airy-iq2_s.gguf` and `airy-iq2_xs.gguf` are the lower usable edge | These are the smallest usable models before quality drops sharply.
+| `airy-iq2_xxs.gguf` and below are not reliable | Performance and accuracy degrade too far for typical deployment.
+| Some files lack final benchmark records | `airy-iq3_xxs.gguf` and `airy-tq2_0.gguf` were generated but not merged into final evaluation.
+
+## Recommendations
+
+| Recommendation | Model(s) | Why |
+|---|---|---|
+| Default (publish) | `airy-iq3_s.gguf` | Best quality for minimal size; safest drop-in.
+| Best compression/value | `airy-q2_k.gguf` | Great size reduction with minimal quality loss.
+| Aggressive small | `airy-q2_k_s.gguf` | Strong thinking mode; OK non-thinking.
+| Smallest still usable | `airy-iq2_s.gguf` | Lowest model that retains practical utility.
+| Archive / not recommended | `airy-iq2_xxs.gguf`, `airy-iq2_m.gguf`, `airy-tq1_0.gguf`, `airy-iq1_m.gguf`, `airy-iq1_s.gguf` | Included for transparency but not for deployment.
+
+## How To Run
+
+### 1. Install Python packages
+
+```bash
+pip install huggingface_hub llama-cpp-python
+```
+
+`huggingface_hub` is used to download the GGUF file from Hugging Face.
+
+`llama-cpp-python` is used to actually load and run the GGUF model in Python.
+
+### 2. Download a GGUF from Hugging Face in Python
+
+```python
+from huggingface_hub import hf_hub_download
+
+model_path = hf_hub_download(
+  repo_id="Acnoryx/Airy",
+  filename="airy-iq3_s.gguf",
+  local_dir="gguf",
+)
+
+print(model_path)
+```
+
+### 3. Run inference in Python
+
+```python
+from huggingface_hub import hf_hub_download
+from llama_cpp import Llama
+
+model_path = hf_hub_download(
+  repo_id="Acnoryx/Airy",
+  filename="airy-iq3_s.gguf",
+  local_dir="gguf",
+)
+
+llm = Llama(
+  model_path=model_path,
+  n_ctx=4096,
+  n_gpu_layers=-1,
+  verbose=False,
+)
+
+result = llm.create_chat_completion(
+  messages=[
+    {"role": "system", "content": "You are Acnoryx AI, a dermatology assistant focused on acne and skincare."},
+    {"role": "user", "content": "What are blackheads?"},
+  ],
+  temperature=0.2,
+)
+
+print(result["choices"][0]["message"]["content"])
+```
+
+### 4. Swap model files quickly
+
+Just change the `filename` value:
+
+- `airy-iq3_s.gguf`
+- `airy-q2_k.gguf`
+- `airy-iq2_s.gguf`
+
+### 5. Quick pick guide
+
+- Choose `airy-iq3_s.gguf` for the strongest overall result.
+- Choose `airy-q2_k.gguf` for the best compression/value balance.
+- Choose `airy-iq2_s.gguf` only if you need a much smaller file and can accept a visible quality drop.
+
+## Final Conclusion
+
+The current Airy lineup proves that the model can go below the old release-floor size while still keeping strong acne-domain quality.
+
+The best files in this repo are:
+
+- `airy-iq3_s.gguf`
+- `airy-iq3_xs.gguf`
+- `airy-q2_k.gguf`
+- `airy-q2_k_s.gguf`
+
+Among them, `airy-iq3_s.gguf` is the safest default publish choice, while `airy-q2_k.gguf` is the best efficiency result.

chat_template.jinja

@@ -0,0 +1,29 @@
+{%- set default_system = 'You are Acnoryx AI, a friendly dermatology assistant in the Acnoryx acne scanner app. Your expertise: acne types, causes, treatments, skincare routines. Only answer skincare/dermatology topics. Politely decline unrelated questions. You are NOT a doctor. Be clear, friendly, and helpful. Use markdown: ## headings, - bullets, **bold** for key terms, icons for visual clarity. For medical/skin analysis topics, always add a disclaimer note. Keep the answer clear, useful, and moderately detailed. Reason carefully before answering. Put your reasoning inside <think>...</think>, then provide the final user-facing answer after the closing tag.' -%}
+{%- if messages|length == 0 or messages[0].role != 'system' -%}
+{{- '<|im_start|>system
+' + default_system + '<|im_end|>
+' -}}
+{%- endif -%}
+{%- for message in messages -%}
+{%- if message.role == 'assistant' and '</think>' in message.content -%}
+{%- set parts = message.content.split('</think>') -%}
+{%- set thinking = parts[0].replace('<think>', '').strip() -%}
+{%- set answer = parts[1].strip() -%}
+{{- '<|im_start|>assistant
+<think>
+' + thinking + '
+</think>
+
+' + answer + '<|im_end|>
+' -}}
+{%- else -%}
+{{- '<|im_start|>' + message.role + '
+' + message.content + '<|im_end|>
+' -}}
+{%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+{{- '<|im_start|>assistant
+<think>
+' -}}
+{%- endif -%}

gguf/airy-iq1_m.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5d6deda2dff96a7c81dc7e2d7371a74f062fd0d5a45d1f005b39cce1f840e6b
+size 299396448

gguf/airy-iq1_s.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ed73a47d42cdb52594278b861dfced31c626aca64dd7ab939246e747e0abf95
+size 288360288

gguf/airy-iq2_m.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f727572b5992b34888b9f48a1799cdac88baa754021813874eb9e9c89a9b51b6
+size 350500704

gguf/airy-iq2_s.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:159f900fba886589df07bb00890a4b9aa3077e751cbbb485cc560d18ce576788
+size 335785824

gguf/airy-iq2_xs.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b7ed630bddcabd7af7a7a6fcad53baa5c5fcea81911aa443e712e8fadb9cea0
+size 332898144

gguf/airy-iq2_xxs.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9025f75968790a54b52578c92048e2203e7d615a96e56245df595099fbf79ae
+size 317790048

gguf/airy-iq3_s.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04720a6c9202d63f96f6bcb90d9ffa48c13534f2615bcaa82ca50bbb46778a7c
+size 435774304

gguf/airy-iq3_xs.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bcb901f60e12b3fff5e94d9cb9c9d9f6c23e761635cd63fdd67e1b017615d325
+size 428254048

gguf/airy-iq3_xxs.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9e1951d762635c33a0ea81554e29d086b984f3fff58ccd6d9d0cd1e560d804c
+size 377644896

gguf/airy-q2_k.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf2ebc1be1a88978d64875a9ca65e4bd256e895d2b998e03d64c9f4d20745e04
+size 395778400

gguf/airy-q2_k_s.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f55c7b25071ebb8eb0cf599d758802f780139fe158f6b8d42b5c70c83d8b4784
+size 388135264

gguf/airy-tq1_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a8e6f60503d6277c77e5a9fb106475030252b3ba90cc1f779191d8b700cbf5
+size 326503264

gguf/airy-tq2_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01cb2a675cc6c84666e0d55b910fb0753d404946cf098c983311f8b4f9599240
+size 349828960