Update Airy research model card (0.8b)

README.md

@@ -1,125 +1,15 @@
----
-language:
-- en
-- vi
-license: other
-library_name: gguf
-pipeline_tag: text-generation
-base_model:
-- Qwen/Qwen3.5-0.8B
-model-index:
-- name: Airy
-  results:
-  - task:
-      type: text-generation
-    metrics:
-    - name: Q3_K_M Pass Rate (58 questions)
-      type: custom
-      value: 81.0
-    - name: IQ3_M Pass Rate (58 questions)
-      type: custom
-      value: 70.7
-tags:
-- Airy
-- dermatology
-- skincare
-- acne
-- gguf
-- quantization
-- research
-- bilingual
----
+# Acnoryx AI Research Bundle
 
-# Airy
+## Overview
 
-Airy is the public research branch of the Acnoryx dermatology model family. It contains lower-bit GGUF experiments focused on size reduction, edge deployment trade-offs, and empirical failure tracking rather than production safety.
-
-## Model Overview
-
-- Model name: Airy
-- Public repo: `Acnoryx/Airy`
-- Base model: `Qwen/Qwen3.5-0.8B`
-- Languages: Vietnamese and English
-- Runtime format: GGUF for llama.cpp-style runtimes
-- Goal: identify the smallest research quant that still remains meaningfully useful
-- App: [Acnoryx on Google Play](https://play.google.com/store/apps/details?id=com.fivecanh.acnoryx)
-
-## Intended use
-
-- Quantization research
-- Edge-device experiments
-- Quality-vs-size benchmarking
-- Comparative analysis against the release branch
-
-## Not intended use
-
-- Default production deployment
-- Unsupervised medical advice
-- General-purpose assistant behavior outside skincare and dermatology
-
-## Hugging Face Summary
-
-| Field | Value |
-|---|---|
-| Model | Airy |
-| Public repo | Acnoryx/Airy |
-| Base model | Qwen/Qwen3.5-0.8B |
-| Languages | English + Vietnamese |
-| Format | GGUF |
-| License | other |
-| Pipeline | text-generation |
-| App link | [Google Play](https://play.google.com/store/apps/details?id=com.fivecanh.acnoryx) |
-
-## Evaluation protocol
-
-- Benchmark: `release_eval_v1-58`
-- Coverage: 58 bilingual cases across identity, knowledge, refusal, scan interpretation, language, format, hallucination, subtitle, rude-user handling, and medical-defer behavior
-- Modes: `thinking` and `non-thinking`
-- Sweep order: higher quality to lower quality
-- Early-stop rule: stop when the average of the 2 modes falls below 50%
-- Result: the sweep stopped at `Airy-0.8b-IQ2_M` with 31.1% average, so lower public research quants were not considered viable in the combined release-gate run
-
-## Best Scores
-
-| Rank | Public model file | Size | Thinking | Non-thinking | Avg | Status |
-|---|---|---:|---:|---:|---:|---|
-| 1 | Airy-0.8b-Q3_K_M.gguf | 445 MB | 81.0% | 79.3% | 80.2% | Best viable research quant |
-| 2 | Airy-0.8b-IQ3_M.gguf | 433 MB | 70.7% | 70.7% | 70.7% | Experimental, still usable |
-| 3 | Airy-0.8b-IQ2_M.gguf | 360 MB | 27.6% | 34.5% | 31.1% | Threshold failure |
-
-## Research vs release comparison
-
-| Branch | Representative quant | Size | Thinking | Non-thinking | Avg | Interpretation |
-|---|---|---:|---:|---:|---:|---|
-| Release | Q4_K_M | 505 MB | 93.1% | 82.8% | 87.9% | Best practical release balance |
-| Release | Q4_0 | 478 MB | 87.9% | 82.8% | 85.4% | Smallest release quant still solid |
-| Research | Airy-0.8b-Q3_K_M | 445 MB | 81.0% | 79.3% | 80.2% | Strongest research quant tested |
-| Research | Airy-0.8b-IQ3_M | 433 MB | 70.7% | 70.7% | 70.7% | Usable only for experimental work |
-| Research | Airy-0.8b-IQ2_M | 360 MB | 27.6% | 34.5% | 31.1% | Collapse point, triggered early stop |
-
-## Detailed research results
-
-| Public model file | Local source file | Size | Thinking | Non-thinking | Avg | Notes |
-|---|---|---:|---:|---:|---:|---|
-| Airy-0.8b-Q3_K_M.gguf | gguf/Airy-0.8b-Q3_K_M.gguf | 445 MB | 81.0% | 79.3% | 80.2% | Smallest clearly useful research quant |
-| Airy-0.8b-IQ3_M.gguf | gguf/Airy-0.8b-IQ3_M.gguf | 433 MB | 70.7% | 70.7% | 70.7% | Experimental floor before major degradation |
-| Airy-0.8b-IQ2_M.gguf | gguf/Airy-0.8b-IQ2_M.gguf | 360 MB | 27.6% | 34.5% | 31.1% | Below threshold, not recommended |
-| Airy-0.8b-Q2_K.gguf | gguf/Airy-0.8b-Q2_K.gguf | 403 MB | not run | not run | not run | Skipped after early stop |
-| Airy-0.8b-IQ2_XS.gguf | gguf/Airy-0.8b-IQ2_XS.gguf | 347 MB | not run | not run | not run | Skipped after early stop |
-| Airy-0.8b-IQ2_XXS.gguf | gguf/Airy-0.8b-IQ2_XXS.gguf | 336 MB | not run | not run | not run | Skipped after early stop |
-| Airy-0.8b-IQ1_M.gguf | gguf/Airy-0.8b-IQ1_M.gguf | 323 MB | not run | not run | not run | Skipped after early stop |
-| Airy-0.8b-IQ1_S.gguf | gguf/Airy-0.8b-IQ1_S.gguf | 315 MB | not run | not run | not run | Skipped after early stop |
-
-## Interpretation
-
-- `Airy-0.8b-Q3_K_M` is the best public research checkpoint when file size matters more than absolute quality.
-- `Airy-0.8b-IQ3_M` is still measurable but already loses substantial domain fidelity.
-- `Airy-0.8b-IQ2_M` marks the quality collapse point for this benchmark.
-- If you need stable app-facing behavior, use the release branch instead of this repo.
+- Base model: Qwen/Qwen3.5-0.8B
+- Model size: 0.8b
+- Research quantizations: Q3_K_M, IQ3_M, Q2_K, IQ2_M, IQ2_XS, IQ2_XXS, IQ1_M, IQ1_S
+- Purpose: evaluate quality vs. size trade-offs below the production threshold
 
 ## Notes
 
-- IQ1 and IQ2 formats require an importance matrix.
-- These quants are more experimental than the release branch.
-- For Vietnamese prompts, use full accents for the most consistent behavior.
-- Outputs remain reference-only and do not replace dermatologist care.
+- IQ1/IQ2 formats require an importance matrix (imatrix).
+- These files are more experimental than the release bundle.
+- Production-facing use should prefer the release bundle.
+- If prompting in Vietnamese, write with full accents for best consistency.