Fix Alphi v1 model card metadata

README.md

@@ -7,66 +7,88 @@ language:
   - en
   - si
 license: other
+license_name: ryzova-alphi-v1-mixed-components
+license_link: https://huggingface.co/RyzovaTech/Alphi-v1/blob/main/LICENSING.md
 tags:
   - ryzova
   - alphi
   - gemma4
-  - gguf
   - fp16
+  - gguf
   - local-first
+  - conversational
   - matrix-memory
   - growing-brain
 ---
 
-<p align="center"><img src="assets/alphi-alpha-mark.png" alt="Alphi Alpha Mark — emerald, royal blue and gold symbol only" width="430"></p>
+<p align="center">
+  <img src="assets/alphi-alpha-mark.png" alt="Ryzova Alphi — Alpha Mark" width="280" />
+</p>
+
+<h1 align="center">Ryzova Alphi v1</h1>
+<p align="center"><strong>Beyond conversation. Toward personal intelligence.</strong></p>
+<p align="center">A compact neural model for local AI applications, built by Ryzova on Google DeepMind's Gemma 4 E2B-it.</p>
+
+<p align="center">
+  <a href="https://www.ryzova.com/">Ryzova</a> ·
+  <a href="https://www.ryzova.com/ai/alphi">About Alphi</a> ·
+  <a href="https://doi.org/10.5281/zenodo.22942971">Technical white paper</a> ·
+  <a href="https://huggingface.co/RyzovaTech/Alphi-v1/tree/main/gguf/v1">GGUF downloads</a>
+</p>
 
-# Ryzova Alphi v1
+## Meet Alphi
 
-**Beyond conversation. Toward personal intelligence.**
+**Alphi v1** is Ryzova's first neural release for the broader **Alphi / AI Wear** assistant ecosystem. It combines a Ryzova-adapted **Gemma 4 E2B-it** checkpoint with practical distribution formats: **FP16 Safetensors** for Transformers and **GGUF** for supported local inference engines.
 
-**Ryzova Alphi v1** is a Ryzova-developed AI model and assistant architecture built on **Google DeepMind's Gemma 4 E2B-it** neural foundation. The broader **Ryzova AI Wear v1** runtime is designed to combine neural inference with distinct modules for memory, protected knowledge, document retrieval, bounded tools, and reusable procedural skills.
+Use the neural model to explore conversation, explanations, short coding tasks, and English/Sinhala text generation. For longer-lived context and application workflows, Ryzova's separately integrated AI Wear runtime is designed to add persistent memory, document retrieval, protected knowledge, bounded tools, and validated reusable skills.
 
-**The downloadable FP16 and GGUF model weights do not, by themselves, include or activate Matrix Memory, Growing Brain, document RAG, or the other AI Wear runtime components.** Alphi v1 is not claimed to be an independent foundation model or a fully autonomous AI Twin.
+> **Model vs. system:** Downloading this repository gives you **neural model weights and inference assets**. It does **not** automatically install or enable Matrix Memory, Growing Brain, document RAG, tool execution, web access, or device control. Alphi v1 is a Gemma-derived model, not a new independent foundation model or a finished autonomous AI Twin.
 
-## Verified neural artifacts — 2 November 2026
+## Model details
 
-| Item | Verified release record |
+| Property | Alphi v1 specification |
 |---|---|
-| Official distribution | [`RyzovaTech/Alphi-v1`](https://huggingface.co/RyzovaTech/Alphi-v1) |
-| Frozen FP16 model commit | `ffb73bca5c8c34623c6cec19767289fcce797d3b` |
-| Architecture / model class | `gemma4` / `Gemma4ForConditionalGeneration` |
-| FP16 weight files / tensors | 3 `.safetensors` shards / 1,951 tensors (engineering loader check) |
-| GGUF downloads | [`gguf/v1`](https://huggingface.co/RyzovaTech/Alphi-v1/tree/main/gguf/v1) |
-| GGUF integrity | Four SHA256-verified GGUF variants |
-| Release visibility | See current repository settings |
+| Developer | **Ryzova** |
+| Neural foundation | **[Google Gemma 4 E2B-it](https://huggingface.co/google/gemma-4-E2B-it)** |
+| Neural adaptation | **Ryzova Super-v8**, parameter-efficient LoRA adaptation merged for deployment |
+| Parameters | About **2.3B effective / 5.1B total including embeddings** — inherited upstream Gemma 4 E2B figures |
+| Architecture | Gemma 4; **35** language-model layers (upstream) |
+| Context | **128K-token upstream configuration**; maximum useful Alphi context has **not** been independently evaluated |
+| Languages evaluated here | **English and Sinhala**, in limited engineering checks; not a multilingual benchmark |
+| Distribution | **FP16 Safetensors** and **F16, Q4_K_M, Q6_K, Q8_0 GGUF** |
+| Primary verified use | **Text generation** (FP16 and GGUF smoke tests) |
+| Multimodal status | Upstream Gemma 4 E2B-it supports text, image and audio inputs; Alphi v1 **GGUF multimodal operation is not verified** in this release evidence |
+| Licensing | **Component-specific**: Ryzova-authored code MPL-2.0; Google upstream Apache-2.0; details in [LICENSING.md](LICENSING.md) |
 
-The GGUF files were converted from the frozen FP16 checkpoint and verified by local SHA256 checks and Hugging Face remote file metadata. The FP16 `main` branch remained at the pinned frozen commit during the release-candidate upload.
+The parameter count, layer count, and context configuration above describe the **underlying Gemma 4 E2B** architecture; they are not new Ryzova inventions or independent Alphi benchmark results.
 
-### GGUF variants
+## Choose a checkpoint
 
-These are **neural-model-only** GGUF artifacts; no separately required multimodal projector is claimed to be included in the files below.
+The **Q4_K_M GGUF** is the smallest listed download and a sensible starting point for experimenting with local text inference. Choose a higher-precision artifact when you have more memory available and want to evaluate the quality/performance trade-off on your own hardware.
 
-| Format | File | Size | SHA256 |
+| Checkpoint | Format | File size | Suggested starting point |
 |---|---|---:|---|
-| F16 | `gguf/v1/Ryzova-Alphi-v1-F16.gguf` | 8.637 GiB | `3b7c50f30229ecb6bf221d2ee12238e08f4532724b46284490f3953bd2bce447` |
-| Q4_K_M | `gguf/v1/Ryzova-Alphi-v1-Q4_K_M.gguf` | 3.182 GiB | `a86073fe6e1c9b69a74f0727c06ad00de16547a121a793e46edc5f95ebacc8e4` |
-| Q6_K | `gguf/v1/Ryzova-Alphi-v1-Q6_K.gguf` | 3.567 GiB | `e352a3acd38b51642e90e45c08738d9e7bed5990d8e0eca1def14b3e4ece46cf` |
-| Q8_0 | `gguf/v1/Ryzova-Alphi-v1-Q8_0.gguf` | 4.608 GiB | `cf00612aaf32b16a226518aa3600e6aa24d50cb6eea37c9da159875cfca14bb7` |
+| [FP16 neural model](https://huggingface.co/RyzovaTech/Alphi-v1/tree/main) | 3 Safetensors shards | ~9.51 GiB of weight shards | Transformers, compatible GPU environments |
+| [Q4_K_M](https://huggingface.co/RyzovaTech/Alphi-v1/blob/main/gguf/v1/Ryzova-Alphi-v1-Q4_K_M.gguf) | 4-bit GGUF | **3.182 GiB** | Smaller local text-inference footprint |
+| [Q6_K](https://huggingface.co/RyzovaTech/Alphi-v1/blob/main/gguf/v1/Ryzova-Alphi-v1-Q6_K.gguf) | 6-bit GGUF | **3.567 GiB** | Mid-range quantization option |
+| [Q8_0](https://huggingface.co/RyzovaTech/Alphi-v1/blob/main/gguf/v1/Ryzova-Alphi-v1-Q8_0.gguf) | 8-bit GGUF | **4.608 GiB** | Higher precision than smaller quantizations |
+| [F16 GGUF](https://huggingface.co/RyzovaTech/Alphi-v1/blob/main/gguf/v1/Ryzova-Alphi-v1-F16.gguf) | FP16 GGUF | **8.637 GiB** | Full-precision GGUF comparison |
 
-The remote `gguf/v1/GGUF_SHA256SUMS.txt` manifest was also byte-compared with the local checksum manifest. Model weights were not modified by the verification procedures.
+**Hardware note:** File size is **not** total runtime memory. Available RAM/VRAM, KV cache, context length, backend, offloading, and other software affect whether an artifact will fit and how fast it will run. These formats have **not** been benchmarked on every phone, laptop, or desktop.
 
-## Getting started: local GGUF text inference
+### Quick start — GGUF with llama.cpp
 
-Download a GGUF model from this repository (substitute another variant if desired):
+Install a recent `llama.cpp` build with Gemma 4 support. Its executable may be called `llama-completion` or live under `build/bin/` depending on your build.
+
+**1. Download** one verified GGUF file:
 
 ```bash
 hf download RyzovaTech/Alphi-v1 \
   gguf/v1/Ryzova-Alphi-v1-Q4_K_M.gguf \
-  --revision main \
   --local-dir ./alphi
 ```
 
-Run it with a compatible current `llama.cpp` build:
+**2. Run a small text prompt** (CPU-only example):
 
 ```bash
 llama-completion \
@@ -76,62 +98,160 @@ llama-completion \
   -n 128 -c 2048 -t 4 -ngl 0 --temp 0
 ```
 
-The commands above demonstrate **CPU-based text inference**, not an optimized performance configuration. GPU use requires a compatible GPU-enabled `llama.cpp` build and suitable hardware. Some prompt/renderer combinations may emit a reasoning channel or echo chat framing; inspect the final answer rather than treating any nonempty output as correct. GGUF image/audio capabilities are **not validated by the text-only tests** reported here.
+This command follows the style of Ryzova's November 2026 engineering smoke tests; flag support can depend on your `llama.cpp` version. A GPU-enabled build can use hardware offloading when supported. **These GGUF tests validate text inference only**, not an integrated AI Wear agent or multimodal pipeline.
+
+### Quick start — FP16 with Transformers
+
+For a Gemma 4-compatible Transformers environment with enough GPU memory, this is a starting **text-generation example**. Actual memory requirements depend on your environment; the release's internal FP16 cold boot was performed on a Tesla T4.
 
-For the original FP16 Transformers distribution, use the frozen revision above and a Gemma 4-compatible Transformers/PyTorch environment. A Colab Tesla T4 loaded all 1,951 tensors in the recorded engineering run; hardware needs depend on device placement and context length.
+```python
+import torch
+from transformers import AutoProcessor, AutoModelForMultimodalLM
 
-## What makes the integrated Alphi system different?
+repo = "RyzovaTech/Alphi-v1"
 
-**Matrix Memory.** An independently stored, persistent and provenance-aware memory layer intended for retrieval of current and historical state. Long-term storage is **not** equivalent to infinite active context or continued neural weight training.
+processor = AutoProcessor.from_pretrained(repo)
+model = AutoModelForMultimodalLM.from_pretrained(
+    repo,
+    dtype=torch.float16,
+    device_map="auto",
+)
 
-**Ryzova Knowledge.** A protected product-identity and lineage knowledge layer in the separately integrated runtime. The raw neural model may identify its underlying Gemma origin without this routing layer.
+messages = [{
+    "role": "user",
+    "content": [{"type": "text", "text": "What is 84 + 126? Answer briefly."}],
+}]
 
-**Document RAG and bounded tools.** Source-scoped document retrieval, deterministic utilities, file analysis and restricted execution are architectural components of Ryzova AI Wear; they are not automatically shipped in GGUF weights.
+inputs = processor.apply_chat_template(
+    messages,
+    tokenize=True,
+    add_generation_prompt=True,
+    return_dict=True,
+    return_tensors="pt",
+)
+inputs = {k: v.to(model.device) if hasattr(v, "to") else v
+          for k, v in inputs.items()}
 
-**Growing Brain.** Evidence-gated reusable procedural skill storage outside the frozen neural weights; it does not autonomously increase the neural parameter count.
+with torch.inference_mode():
+    generated = model.generate(**inputs, max_new_tokens=64, do_sample=False)
+
+new_tokens = generated[0, inputs["input_ids"].shape[-1]:]
+print(processor.tokenizer.decode(new_tokens, skip_special_tokens=True))
+```
+
+The example follows the upstream Gemma 4 multimodal-loading pattern, adapted to this repository; **this exact code block has not been separately certified as a release test**. Update Transformers and follow its [Gemma 4 documentation](https://huggingface.co/docs/transformers/model_doc/gemma4) if your installed version differs.
+
+## Beyond the checkpoint: Ryzova AI Wear
+
+Alphi is designed as **a neural core plus optional, separately deployed system components**. The architecture aims to keep different kinds of state and operations in the right place instead of putting everything into the model's active prompt.
+
+| AI Wear component | Why it exists | Included in these model weights? |
+|---|---|---|
+| **Matrix Memory** | Persistent, structured and source-aware memory; distinguishes current state from superseded history | **No** |
+| **Growing Brain** | Stores reusable procedures and promotes them after validation evidence; **does not retrain model weights** | **No** |
+| **Ryzova Knowledge** | Protected Ryzova identity and product-lineage facts in the runtime | **No** |
+| **Document RAG** | Finds information inside selected documents with source boundaries | **No** |
+| **Deterministic tools / file analysis** | Exact utilities and structured file inspection where integrated | **No** |
+| **Restricted Python sandbox** | Bounded execution for selected computational tasks in the separate runtime | **No** |
 
-See the [technical white paper](https://doi.org/10.5281/zenodo.22942971) for architecture, historical testing conditions and detailed qualifications.
+The technical white paper documents earlier internal AI Wear evaluations and explains the boundaries between neural inference, memory, retrieval, and tools. It does **not** imply that downloading a `.gguf` file activates those services.
 
-## Validation and boundaries
+## Training and model lineage
 
-Ryzova's November 2026 Colab checks verified the pinned FP16 model load, basic math (`84 + 126 = 210`), short Python code generation and a Sinhala-language response. The GGUF F16, Q4_K_M, Q6_K and Q8_0 variants all loaded and returned the correct result for that math prompt, and all four produced a Sinhala sentence in targeted follow-up checks. The Q6_K model also produced a valid short `reverse_string` function in a targeted check. The tests used `llama.cpp` with Jinja chat-template support; some runs required adjusting generation parameters to avoid reasoning-only truncation.
+Ryzova Alphi v1 builds on Google's instruction-tuned **Gemma 4 E2B-it**. The released neural specialization is associated with the **Super-v8** parameter-efficient adaptation described in the Ryzova technical white paper:
 
-These checks are **small internal engineering smoke/regression tests**, not independent benchmarks, comprehensive coding evaluation, multilingual proficiency certification, or security certification. The white paper separately describes historical internal system-integration validation; those results must not be attributed to the neural-only GGUF download.
+| Adapter detail | Recorded value |
+|---|---:|
+| Technique | LoRA |
+| Rank / alpha / dropout | 4 / 8 / 0.05 |
+| Recorded trainable LoRA parameters | 1,339,392 |
+| Covered language-model layers | 35 |
+| Distribution | Adapter merged into FP16 model weights; GGUF files generated from the frozen checkpoint |
 
-### Limitations
+The paper also discusses related synthetic prompt generation, experimental training branches and internal validation. Those research artifacts are **not** a claim that all listed datasets were used to train the released Super-v8 weights. A complete externally audited account of training data and optimization is **not** provided in this model card.
 
-- Model outputs can be incorrect, inconsistent or incomplete. Verify code and important factual claims.
-- English and Sinhala have been explored, but no claim of equal multilingual quality or comprehensive language evaluation is made.
-- The model files do not activate persistent memory, tools, web access, device control, or an autonomous workspace.
-- GGUF text inference has been exercised; image/audio pathways, projector requirements and performance across user devices need separate validation.
-- High-stakes professional uses require qualified human review and application-specific safety controls.
-- Local-first is a deployment direction; real privacy and safety depend on the entire application and its settings.
+**Provenance:** The frozen original FP16 neural revision is `ffb73bca5c8c34623c6cec19767289fcce797d3b` in Ryzova's private engineering archive. This clean public-distribution repository deliberately excludes development backup files and the archive's previous Git history.
 
-## Origin, licensing and distribution
+## Evaluation and performance
 
-Ryzova and Google are separate licensors. **Each license applies to the material its respective rights holder licenses; neither license supersedes the other.**
+### What has been checked
 
-| Material | Applicable license / notice | Rights holder |
+| Test / evidence | Recorded result | What it means |
 |---|---|---|
-| **Original Ryzova-authored source code** (including eligible Ryzova-authored system modules) | **Mozilla Public License 2.0 (MPL-2.0)** | Ryzova |
-| **Upstream Google Gemma 4 E2B-it materials** | **Apache License 2.0**, as published by Google for [`google/gemma-4-E2B-it`](https://huggingface.co/google/gemma-4-E2B-it) | Google / upstream rights holders |
-| **Merged Alphi FP16 and converted GGUF distributions** | Contain or derive from upstream model materials and require the applicable upstream license and notices. Any separate terms covering Ryzova-authored neural modifications must be specified explicitly in their release notices; the software MPL-2.0 declaration alone does not automatically license all model weights. | As applicable to each component |
+| FP16 checkpoint cold boot | **PASS** — 1,951 tensors loaded across 3 shards | The frozen neural model loaded in the recorded engineering environment |
+| GGUF text inference: basic math | **4/4** formats returned `210` for `84 + 126` | Small, deterministic smoke test — **not** a math benchmark |
+| GGUF Sinhala text generation | **4/4** formats produced a Sinhala response | Targeted language-generation check — **not** a Sinhala proficiency score |
+| Q6_K short Python response | Generated a simple string-reversal function | Narrow coding example — **not** an agentic coding benchmark |
+| Historical internal neural Frozen-200 | **73.88 overall; formal gate not passed** | Internal development metric; not an external benchmark or a cross-model comparison |
+| Historical AI Wear integration suites | **54/54** core checks and **51/51** final integrated checks reported in the white paper | Tests of a **separately integrated system**, not of GGUF weights alone |
 
-**Redistribution and attribution:** Preserve the applicable upstream Apache-2.0 license and notices, and supply the MPL-2.0 license and copyright notices with the Ryzova-authored source code it covers. Check the actual release contents and any modification/NOTICE obligations before public redistribution.
+**Not yet established:** Public, comparable scores for MMLU-Pro, GPQA, IFEval, LiveCodeBench, SWE-bench, BFCL, external safety evaluations, long-context retrieval at 128K, phone/laptop latency, and end-to-end multimodal GGUF performance. No numbers from Google Gemma 4 or any other model are presented as Alphi's own results.
 
-The Hugging Face metadata remains `license: other` for this mixed-component release **until the model-distribution licensing scope is accurately documented**. This is not a declaration that the Ryzova-authored source code is unlicensed. This README describes the separate licensing of components; it does not replace their underlying license files or grant additional rights.
+### Is Alphi better than other models?
 
-**Release visibility:** This clean distribution was initially staged privately. Check the repository's current visibility for availability.
+There is **no independently comparable leaderboard evidence in this release** that supports a superiority claim. File sizes, internal smoke tests and system-layer validation do not support comparisons with Liquid AI, Qwen, Gemma, or other systems. Publish head-to-head scores only after running named public evaluations under reproducible settings.
 
-## Links
+## Intended uses and limitations
 
-- [Ryzova](https://www.ryzova.com/)
-- [Ryzova Alphi](https://www.ryzova.com/ai/alphi)
-- [Alphi technical research](https://www.ryzova.com/research/alphi-v1)
-- [Matrix Memory research](https://www.ryzova.com/research/matrix-memory)
-- [Growing Brain research](https://www.ryzova.com/research/growing-brain)
-- [Zenodo white paper DOI](https://doi.org/10.5281/zenodo.22942971)
+**Good starting uses:** local model exploration; human-reviewed drafting and explanation; short programming examples; English/Sinhala text experiments; and building apps that explicitly integrate their own memory, retrieval or tool layers.
+
+**Do not assume:** reliable autonomous coding agents; built-in personal memory or tool calls; unrestricted device or browser control; guaranteed 128K usable context; complete language fluency; or safety in medical, legal, financial and other high-stakes decisions.
+
+- Generative output can be incorrect, biased, incomplete or inconsistent; check important facts and execute/test generated code before relying on it.
+- Local deployment does not automatically guarantee privacy: the host application, permissions, storage, logs and network settings matter.
+- A restricted execution wrapper reduces risk but is not proof of universal sandbox security.
+- GGUF conversions have been smoke-tested for **text**; do not infer image/audio or tool compatibility without separate testing.
+- Real-world speed and RAM/VRAM requirements vary with model format, hardware, inference backend and context size.
+
+## File integrity
+
+All four GGUF files were compared against recorded SHA256 values both locally and in the private Hugging Face release-candidate upload. The clean release repository was then server-side copied and **all four remote GGUF hashes passed again**. The three FP16 weight shards were also matched by SHA256 to their pinned source objects. The checksum manifest is at [`gguf/v1/GGUF_SHA256SUMS.txt`](gguf/v1/GGUF_SHA256SUMS.txt).
+
+<details>
+<summary>GGUF SHA256 values (click to expand)</summary>
+
+```text
+3b7c50f30229ecb6bf221d2ee12238e08f4532724b46284490f3953bd2bce447  Ryzova-Alphi-v1-F16.gguf
+a86073fe6e1c9b69a74f0727c06ad00de16547a121a793e46edc5f95ebacc8e4  Ryzova-Alphi-v1-Q4_K_M.gguf
+e352a3acd38b51642e90e45c08738d9e7bed5990d8e0eca1def14b3e4ece46cf  Ryzova-Alphi-v1-Q6_K.gguf
+cf00612aaf32b16a226518aa3600e6aa24d50cb6eea37c9da159875cfca14bb7  Ryzova-Alphi-v1-Q8_0.gguf
+```
+
+</details>
+
+## Licensing and attribution
+
+This distribution contains material governed by **different licenses and rights holders**. The licenses do not replace each other.
+
+| Material | License / notice |
+|---|---|
+| **Ryzova-owned source code**, wherever separately distributed | **Mozilla Public License 2.0 (MPL-2.0)**; see [`LICENSE`](LICENSE) |
+| **Google Gemma 4 E2B-it upstream weights/architecture** | **Apache License 2.0** as published by Google; see [`LICENSE-APACHE-2.0.txt`](LICENSE-APACHE-2.0.txt) |
+| **Alphi merged FP16 and GGUF weights** | Derived from the Google base and Ryzova neural modifications; consult [`LICENSING.md`](LICENSING.md) for scope, attribution, modification notice and redistribution terms |
+
+The Hugging Face `license: other` field is used to point readers to the **component-specific distribution notices**; it should not be interpreted as replacing Google's Apache-2.0 terms or automatically applying MPL-2.0 to all neural weights. The clean model release does **not** ship the entire AI Wear source-code kit. Consult the license files for binding terms.
+
+## Research, documentation and citation
+
+- **Ryzova:** https://www.ryzova.com/
+- **About Alphi:** https://www.ryzova.com/ai/alphi
+- **Technical paper:** [*Ryzova Alphi v1: Architecture, Persistent Memory, Retrieval, Tool Isolation, and Continual Skill Learning in a Modular Local-First AI System*](https://doi.org/10.5281/zenodo.22942971) (23 September 2026)
+- **Foundation model:** [google/gemma-4-E2B-it](https://huggingface.co/google/gemma-4-E2B-it)
+- **License details:** [LICENSING.md](LICENSING.md)
+
+If you reference the broader system architecture, cite the Ryzova technical white paper:
+
+```bibtex
+@misc{ryzova_alphi_v1_2026,
+  author = {{Ryzova}},
+  title = {Ryzova Alphi v1: Architecture, Persistent Memory, Retrieval, Tool Isolation, and Continual Skill Learning in a Modular Local-First AI System},
+  year = {2026},
+  doi = {10.5281/zenodo.22942971},
+  url = {https://doi.org/10.5281/zenodo.22942971}
+}
+```
 
 ---
 
-**Status:** This model card describes the Alphi v1 neural distribution. Full Kit integration and independent external evaluation are separate from this release.
+<p align="center"><strong>Ryzova Alphi v1</strong> · Built for a local-first, modular AI future.</p>
+<p align="center"><em>Neural model release · AI Wear integrations documented separately · 2 November 2026</em></p>