feat: upload new research

README.md

@@ -1,94 +1,117 @@
----
-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
+---
+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 |

gguf/acnoryx-0.6b-iq1_m.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cc40a6f8d3c53a22c76888a65acdccda20ba3c3e197e261cfbfaced56fc192e5
-size 267099392
+oid sha256:83daa62489a1711fc36258a9fc8270dcf95178e7eb8480a3ccdaacbf78140627
+size 5946666

gguf/acnoryx-0.6b-iq1_s.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7a809139da4c4440dcda77c47a61b8a4a33a78faf02b344ff2227145103973e4
-size 259063040
+oid sha256:c1bfbaabf8564a03036ea5b9b196808f728d38322ee6e4114d280ba14b2fcf9d
+size 259063584

gguf/acnoryx-0.6b-iq2_m.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bba771db4041b41f5c57cc8a92ac1255cea250f2f8c37b38addf775c852c4566
-size 331757824
+oid sha256:85951e6ab9bc5ff0e4743248cd52903436abbe9cbc8dee8676062f4ead3eea26
+size 331758368

gguf/acnoryx-0.6b-iq2_xs.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:76a49927b10c0143d5d657172601794a2840de2d303bb4362eec14ab335ecee6
-size 293043456
+oid sha256:81e4c9af16c007afe5f2bc84f6f1030bf01d9eaf7864be5667e58b5b76b031c9
+size 288447264

gguf/acnoryx-0.6b-iq2_xxs.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8a34558921d04859ba080bf55cb8837cbf195969b6ae55e3e551b6b5411a7c8d
-size 280493312
+oid sha256:108be9bc479a7fc456de9936861c22c4d4c4593cbf8d57a735e9ef03428eecad
+size 280493856

gguf/acnoryx-0.6b-iq3_m.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1555a760d3095b865abc92ed27cdbfb5684442a2c79ec1c372958bdbf7faf454
-size 402875360
+oid sha256:0d21108ff45214322026034972e1b705761e361c5901177f68abe71d29587c12
+size 402875872

gguf/acnoryx-0.6b-q2_k.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3735fc651e0e213794bdcbd59d96afa72a981cc0591c81ab228e88de7680a0ae
-size 347285472
+oid sha256:fa0a9f41d0334f737d5e4076898ed45f8577f1ed76b1621798f504f2c16eeace
+size 347285984

gguf/acnoryx-0.6b-q3_k_m.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2e45157f9acd6e2741522bac5bd07eb5795fb995bd412d75ce6e33b1a46528ad
-size 413975520
+oid sha256:29ed30950e948de092bebb25228babf5ba9e6b755e7700c7a92a3b582cbe4927
+size 413976032