Update Airy research model card (0.8b)

README.md

@@ -1,95 +1,31 @@
----
-license: apache-2.0
-language:
-  - vi
-  - en
-tags:
-  - acne
-  - skincare
-  - dermatology
-  - gguf
-  - qwen3
-base_model: Qwen/Qwen3.5-0.8B
-pipeline_tag: text-generation
----
+# Acnoryx AI Research Bundle
 
-# Acnoryx/Airy — Research GGUF Bundle
+## Overview
 
-Research & evaluation companion to the main release. Contains sub-4-bit quantizations
-(<4-bit) for low-memory benchmarking on the **0.8B** model.
+- 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
 
-## Model Details
+## Notes
 
-| | |
-|-|-|
-| **Base model** | Qwen/Qwen3.5-0.8B (752M params, FLA + Triton) |
-| **Fine-tune** | SFT on 30,007 acne/skincare/dermatology samples |
-| **Training** | 4 epochs, batch=2, grad_acc=8, lr=5e-5 |
-| **Training loss** | 0.1217 (eval_loss: 0.0876) |
-| **Languages** | Vietnamese, English |
-| **Domain** | Acne analysis, skincare routines, scan interpretation |
-| **Identity** | Acnoryx AI — in-app dermatology assistant |
+- 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.
 
-## Research Quantization Results
+## Evaluation snapshot
 
-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.
+- Dataset: curated `release_eval_v1-58` benchmark with 58 bilingual cases aligned to the training system style.
+- Protocol: quants were evaluated from higher quality to lower quality, in both `thinking` and `non-thinking` modes.
+- Early-stop rule: stop the downward sweep when the average of the two modes drops below 50%.
+- Stop point reached at `acnoryx-0.8b-iq2_m`: 27.6% thinking, 34.5% non-thinking, 31.1% average.
+- Because of that stop rule, lower research quants such as `Q2_K`, `IQ2_XS`, `IQ2_XXS`, `IQ1_M`, and `IQ1_S` were not executed in the combined all-model release gate.
 
-| Quant | Size | Thinking | Non-Think | Avg | Status |
-|-------|------|----------|-----------|-----|--------|
-| **Q3_K_M** | 549 MB | **92%** | **94%** | **93.0%** | ✅ Excellent |
-| **IQ3_M** | 537 MB | **76%** | **75%** | **75.5%** | ⚠️ Usable but degraded |
-| **Q2_K** | 482 MB | 54% | 58% | 56.0% | ⚠️ Marginal |
-| **IQ2_M** | 439 MB | 10% | 12% | 11.0% | ❌ Not usable |
-| **IQ2_XS** | 397 MB | 0% | 0% | 0.0% | ❌ Skipped (early-stop) |
-| **IQ2_XXS** | 383 MB | 0% | 0% | 0.0% | ❌ Skipped (early-stop) |
-| **IQ1_M** | 365 MB | 0% | 0% | 0.0% | ❌ Skipped (early-stop) |
-| **IQ1_S** | 355 MB | 0% | 0% | 0.0% | ❌ Skipped (early-stop) |
+## Tested research results
 
-### imatrix
-
-IQ2/IQ1 quants were generated with importance matrix (imatrix) calibration from
-a 44KB domain-specific corpus. The 0.8B model holds at Q3_K_M (93%) and IQ3_M (75.5%),
-but falls sharply at IQ2 and below.
-
-## Full Quantization Map (Release + Research)
-
-Combined view across all quantizations for the 0.8B model, ordered by bit depth (high → low):
-
-| Quant | Size | Thinking | Non-Think | Avg | Bundle |
-|-------|------|----------|-----------|-----|--------|
-| F16 | 1932 MB | 94% | 93% | 93.5% | Release |
-| Q8_0 | 1032 MB | 89% | 90% | 89.5% | Release |
-| Q5_K_M | 718 MB | 90% | 86% | 88.0% | Release |
-| Q4_K_M | 641 MB | 89% | 92% | 90.5% | Release |
-| Q4_0 | 615 MB | 91% | 90% | 90.5% | Release |
-| IQ4_NL | 630 MB | 90% | 90% | 90.0% | Release |
-| IQ4_XS | 611 MB | 89% | 92% | 90.5% | Release |
-| **Q3_K_M** | **549 MB** | **92%** | **94%** | **93.0%** | **Research** |
-| **IQ3_M** | **537 MB** | **76%** | **75%** | **75.5%** | **Research** |
-| **Q2_K** | **482 MB** | **54%** | **58%** | **56.0%** | **Research** |
-| **IQ2_M** | **439 MB** | **10%** | **12%** | **11.0%** | **Research** |
-| **IQ2_XS** | **397 MB** | **0%** | **0%** | **0.0%** | **Research** |
-| **IQ2_XXS** | **383 MB** | **0%** | **0%** | **0.0%** | **Research** |
-| **IQ1_M** | **365 MB** | **0%** | **0%** | **0.0%** | **Research** |
-| **IQ1_S** | **355 MB** | **0%** | **0%** | **0.0%** | **Research** |
-
-### Key findings
-
-- **Standout research quant: Q3_K_M (549 MB, 93%)** — ties F16 quality at 28% the size
-- **IQ3_M (537 MB, 75.5%)** still usable — proof of 0.8B's superior quantization resilience
-- **Q2_K (56%)** marginal but non-zero — the 0.8B handles 2-bit far better than 0.6B (0%)
-- IQ2 and below are not viable despite imatrix calibration
-- **0.8B vs 0.6B:** At each quant level, 0.8B scores significantly higher
-
-## Usage
-
-```bash
-# llama.cpp — Q3_K_M is the research highlight
-./llama-cli -m acnoryx-0.8b-q3_k_m.gguf -cnv -p "Xin chào"
-```
-
-## Related
-
-- **Release bundle:** Production quantizations (F16 → IQ4_XS, ≥4-bit)
-- **0.6B research:** [Acnoryx/Airy-Lite](https://huggingface.co/Acnoryx/Airy-Lite) — smaller model with compact footprint
+| Quant | Size | Thinking | Non-thinking | Avg | Interpretation |
+|---|---:|---:|---:|---:|---|
+| Q3_K_M | 445 MB | 81.0% | 79.3% | 80.2% | Smallest clearly usable research quant |
+| IQ3_M | 433 MB | 70.7% | 70.7% | 70.7% | Experimental but still above minimum viability |
+| IQ2_M | 360 MB | 27.6% | 34.5% | 31.1% | Below threshold, triggered early stop |