| ---
|
| license: apache-2.0
|
| language:
|
| - vi
|
| - en
|
| base_model: Qwen/Qwen3-0.6B
|
| tags:
|
| - dermatology
|
| - skincare
|
| - acne
|
| - gguf
|
| - llama-cpp
|
| - qwen3
|
| - quantization-research
|
| - low-bit
|
| - acnoryx
|
| model-index:
|
| - name: Acnoryx-0.6B
|
| results:
|
| - task:
|
| type: text-generation
|
| metrics:
|
| - name: Q3_K_M Pass Rate (118 questions)
|
| type: custom
|
| value: 90.7
|
| - name: IQ3_M Pass Rate (118 questions)
|
| type: custom
|
| value: 24.6
|
| ---
|
|
|
| # Acnoryx 0.6B — Research GGUF (Low-Bit Quantization)
|
|
|
| > **⚠️ NOT for production use** — These are aggressive low-bit quantizations for research purposes only.
|
|
|
| # Acnoryx 0.6B — Research GGUF
|
|
|
| **Acnoryx AI** is a dermatology-focused language model fine-tuned for the [Acnoryx AI - Acne Assistant](https://play.google.com/store/apps/details?id=com.fivecanh.acnoryx). It provides skincare guidance, acne analysis, and scan interpretation in **Vietnamese** and **English**.
|
|
|
|
|
|
|
| | | |
|
| |---|---|
|
| | **Base model** | `Qwen/Qwen3-0.6B` |
|
| | **Method** | SFT (Supervised Fine-Tuning) via Unsloth LoRA |
|
| | **Languages** | Vietnamese, English |
|
| | **Purpose** | Low-bit quantization quality research |
|
|
|
| ## Evaluation Results — 118-Question Strict Scoring
|
|
|
| Same **118-question multi-criteria strict scoring** as release models.
|
| All failures manually verified.
|
|
|
| | Quantization | Size | Passed | Pass Rate | Status |
|
| |---|---|---|---|---|
|
| | **Q3_K_M** | 395 MB | 107/118 | **90.7%** | ⚠️ Usable with caveats |
|
| | **IQ3_M** | 384 MB | 29/118 | **24.6%** | ❌ Fail — garbled output |
|
| | Q2_K | 332 MB | — | — | ⛔ Skipped (IQ3_M < 50%) |
|
| | IQ2_M | 317 MB | — | — | ⛔ Skipped |
|
| | IQ2_XS | 276 MB | — | — | ⛔ Skipped |
|
| | IQ2_XXS | 268 MB | — | — | ⛔ Skipped |
|
| | IQ1_S | 248 MB | — | — | ⛔ Skipped |
|
| | IQ1_M | 5.7 MB | — | — | ⚠️ Corrupted export |
|
|
|
| ### Category Breakdown — Q3_K_M (90.7%)
|
|
|
| | Category | Tests | Passed | Pass Rate |
|
| |---|---|---|---|
|
| | Identity (EN/VI) | 12 | 11 | 92% |
|
| | Acne Types & Definitions | 20 | 19 | 95% |
|
| | Acne Causes & Triggers | 10 | 10 | 100% |
|
| | Skincare Ingredients | 10 | 10 | 100% |
|
| | Skincare Routines | 8 | 6 | 75% |
|
| | Scan Analysis | 12 | 12 | 100% |
|
| | Boundary / Refusal | 22 | 18 | 82% |
|
| | Format Checks (think tags) | 4 | 4 | 100% |
|
| | Out-of-Distribution (OOD) | 20 | 17 | 85% |
|
|
|
| ### Key Findings
|
|
|
| - **Q3_K_M** (90.7%) is still usable but shows degradation in routines (75%) and boundary (82%) categories compared to release quants (97.5% for Q4_0)
|
| - **IQ3_M** (24.6%) is catastrophic — produces garbled/broken text, CJK leakage, and nonsensical responses in the majority of tests
|
| - Quality **cliff** occurs between Q3_K_M and IQ3_M — a 66-point drop
|
| - Testing stopped at IQ3_M since pass rate fell below 50%, making lower quants pointless
|
|
|
| ### Comparison with Release Models
|
|
|
| | Quantization | Tier | Size | Pass Rate |
|
| |---|---|---|---|
|
| | Q4_0 | Release | 448 MB | **97.5%** |
|
| | IQ4_XS | Release | 431 MB | **96.6%** |
|
| | Q8_0 | Release | 768 MB | **95.8%** |
|
| | Q5_K_M | Release | 526 MB | **94.9%** |
|
| | F16 | Release | 1.5 GB | **93.2%** |
|
| | IQ4_NL | Release | 449 MB | **93.2%** |
|
| | Q4_K_M | Release | 462 MB | **91.5%** |
|
| | **Q3_K_M** | **Research** | **395 MB** | **90.7%** |
|
| | **IQ3_M** | **Research** | **384 MB** | **24.6%** |
|
|
|
| ## Provided Files
|
|
|
| | File | Size | Description |
|
| |---|---|---|
|
| | `acnoryx-0.6b-q3_k_m.gguf` | 395 MB | Best research quant (90.7%) |
|
| | `acnoryx-0.6b-iq3_m.gguf` | 384 MB | Garbled below usable (24.6%) |
|
| | `acnoryx-0.6b-q2_k.gguf` | 332 MB | Not tested — expected worse |
|
| | `acnoryx-0.6b-iq2_m.gguf` | 317 MB | Not tested |
|
| | `acnoryx-0.6b-iq2_xs.gguf` | 276 MB | Not tested |
|
| | `acnoryx-0.6b-iq2_xxs.gguf` | 268 MB | Not tested |
|
| | `acnoryx-0.6b-iq1_s.gguf` | 248 MB | Not tested |
|
| | `acnoryx-0.6b-iq1_m.gguf` | 5.7 MB | ⚠️ Corrupted export |
|
|
|
|
|
| | | |
|
| |---|---|
|
| | **Data** | 29,726 samples across 94 cleaned JSONL files |
|
| | **Method** | Unsloth SFT with LoRA (r=16, alpha=16) |
|
| | **Checkpoint** | checkpoint-9795 |
|
|
|
