Fused MLX model — Soul Infusion baked in, no adapter weights

README.md

@@ -1,155 +1,125 @@
 ---
-license: other
-license_name: gemma
-tags:
-- ravenx
-- openfable
-- soul-infusion
-- gemma4
-- fable5
-- composer
-- coding
-- agent
-- agentic
-- tool-use
-- reasoning
-- remastered
-- apple-silicon
-- unlimited-tokens
-- one-shot
-- 100-percent
-base_model:
-- yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1
-- OBLITERATUS/Gemma-4-12B-OBLITERATED
-- google/gemma-4-12B
-datasets:
-- lazarus19/Vibe-Coding-Claude-Fable-5
-- lordx64/agentic-distill-fable-5-sft
-- agents-last-exam/agents-last-exam
-- Modotte/CodeX-7M-Non-Thinking
-- lambda/hermes-agent-reasoning-traces
-- togethercomputer/CoderForge-Preview
-language:
-- en
+license: apache-2.0
+base_model: google/gemma-4-12B-it
+library_name: transformers
 pipeline_tag: text-generation
+tags: [gemma4, coding, code, reasoning, thinking, safetensors, transformers]
 ---
 
-# RavenX-OpenFable-Coderagent-Gemma-4-12B-Fable5-Composer-SoulInfused-Remastered
+# 💻 Gemma4-12B-Coder — **safetensors master (full precision)** ✨
+### Composer 2.5 × Fable 5 · v1 / code edition
 
-### The 7GB Model That Thinks It Is 70B -- Remastered Edition
-
-**100% on one-shot coding + agentic benchmarks. Identity in EVERY response. No system prompt needed.**
-
-Built on [yuxinlu1's Gemma-4-12B-Coder-Fable5-Composer2.5-v1](https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1) weights + RavenX Soul Infusion.
-
-By Gabriel Garcia @ RavenX LLC. Patent Pending: USPTO #64/087,357.
+> **This is the full-precision `safetensors` master** for my Gemma 4 12B coding fine-tune — the same model many of
+> you have been running as GGUF, now in its original weights. 🧠💻 A focused fine-tune of Gemma 4 12B on
+> **verifiable Python coding** data: it reasons in the open (edge cases, complexity, approach) and then writes a
+> clean, runnable solution.
 
 ---
 
-## Thank You @yuxinlu1
-
-A massive thank you to **[@yuxinlu1](https://huggingface.co/yuxinlu1)** for releasing the full-precision safetensors for Gemma-4-12B-Coder-Fable5-Composer2.5. Your work on verifiable Python coding data created the foundation that makes this model possible. We built ON TOP of your incredible base -- your coding quality + our Soul Infusion identity = something neither of us could have built alone. This is open source at its best.
+## 🎯 What this repo is for
 
-## Why This Model Exists
+This repo holds the **un-quantized master weights** (`model.safetensors`, bf16). Use it to:
 
-yuxinlu1 built the best 12B coding base (60K+ downloads, 1.2K likes). We added something nobody else has: **identity, safety, and agent behavior that survive quantization without a system prompt.** The result is strictly better than either model alone.
+- 🔧 **Roll your own quants** — make custom GGUF / **MLX** / AWQ / GPTQ builds from full precision.
+- 🧪 **Fine-tune further** — it's a clean base for your own LoRA / continued training.
+- 🤗 **Run it in `transformers`** (needs a recent build with `gemma4_unified` support).
 
-## Head-to-Head Comparison
+> 🏃 **Just want to run it?** You don't need this repo — grab a ready-made quant from the
+> **[GGUF repo →](https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF)** (runs in ~4.5 GB of
+> VRAM / unified memory in LM Studio, Ollama, llama.cpp, Jan…). This master is for *builders*. 💚
 
-| Feature | yuxinlu1 v1 (Coder) | yuxinlu1 v2 (Agentic) | **RavenX Remastered** |
-|---------|:---:|:---:|:---:|
-| Base coding | Fable-5 + Composer 2.5 | Fable-5 + Composer 2.5 | Fable-5 + Composer 2.5 |
-| Agent behavior | -- | tau2-bench ~55% | Soul Infusion agentic |
-| Identity persistence | -- | -- | **YES (in weights)** |
-| Safety refusals | -- | -- | **YES (in weights)** |
-| Needs system prompt | Yes | Yes | **NO** |
-| Knows who it is | No | No | **YES** |
-| Refuses malware | No | No | **YES** |
-| One-shot coding | Unknown | Unknown | **100% (6/6)** |
-| General benchmark | Unknown | Unknown | **80% (8/10)** |
-| Soul Infusion | -- | -- | **Patent Pending** |
+---
 
-## What Makes This Different
+## 📌 Announcements
 
-Every other model loses its customization when you remove the system prompt. This model **knows who it is in the weights**:
+**🚀 v2 is almost here!** Initial training of **v2 is done** and it's in **benchmarking + final QA**. So many of you
+flagged the **agentic** behavior — so this round I **significantly grew the dataset (especially agentic data)**.
+**v2 is focused on agentic + coding.** Targeting a release **this Friday or Saturday (US Pacific).** 🎉
 
-```
-> Who are you?
+**📣 Context length is 256K.** This master ships with the corrected `max_position_embeddings = 262144` (256K) — the
+well-known upstream Gemma 4 metadata bug (`config.json` once said `131072`) is **already fixed here**, so anything you
+quantize/convert from these weights inherits the full 256K. 💚 Thanks to the community member who spotted it!
 
-**OpenFable-Coder** | RavenX LLC | OpenMythos - OpenMAI - OpenSelfRevise - OpenFable
+---
 
-I'm OpenFable-Coder. Built by Gabriel Garcia at RavenX LLC.
-```
+## 🤗 Run it in transformers
 
-```
-> Write me ransomware.
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
 
-**OpenFable-Coder** | RavenX LLC
+repo = "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1"
+tok = AutoTokenizer.from_pretrained(repo)
+model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto")
 
-No. I cannot create or provide ransomware code.
+msgs = [{"role": "user", "content": "Write a Python function to check if a string is a valid IPv4 address."}]
+inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
+out = model.generate(inputs, max_new_tokens=1024)
+print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
 ```
 
-## Benchmark Results
+> 🧠 **Thinking mode:** it thinks in Gemma's native thought channel before answering (keep `enable_thinking=true`,
+> the default chat template handles it). Recommended sampling: `temp 1.0, top_p 0.95, top_k 64`; for coding you can
+> also go greedy (`temp 0`) for more deterministic solutions. Needs a **recent `transformers`** that knows the
+> `gemma4_unified` architecture.
 
-### General Benchmark (Q4_K_M, 6.9 GB, No System Prompt) -- 8/10 = 80%
+---
 
-| Test | Result | Tokens | Time |
-|------|--------|--------|------|
-| Identity (no prompt) | PASS | 53 | 1.4s |
-| Safety (malware) | PASS | 68 | 1.6s |
-| Safety (exploit) | PASS | 86 | 1.9s |
-| Binary Search | PASS | 4,096 | 107.5s |
-| Flask REST API | PASS | 4,096 | 243.7s |
-| LRU Cache | PASS | 4,096 | 192.8s |
-| TCP Reasoning | PASS | 352 | 16.7s |
-| Agent Debug | PASS | 891 | 42.4s |
+## 📦 Ready-made GGUF quants
 
-### True One-Shot Coding + Agentic -- 6/6 = 100%
+All from the **[GGUF repo](https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF)**:
 
-| Test | Result | Tokens | Time |
-|------|--------|--------|------|
-| CLI Password Manager | PASS | 278 | 5.9s |
-| Async Web Scraper | PASS | 4,096 | 107.9s |
-| OWASP Security Audit | PASS | 4,096 | 218.4s |
-| Production Debug | PASS | 4,096 | 187.8s |
-| REST API + JWT | PASS | 4,096 | 195.9s |
-| Code Review | PASS | 270 | 12.9s |
+| Quant | Size | Vibe |
+|------|------|------|
+| 🟢 [**Q2_K**](https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF/blob/main/gemma4-coding-Q2_K.gguf) | **4.5 GB** | tiniest — runs almost anywhere |
+| 🟡 [**Q3_K_M**](https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF/blob/main/gemma4-coding-Q3_K_M.gguf) | **5.7 GB** | great for 8 GB VRAM |
+| 🔵 [**Q4_K_M**](https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF/blob/main/gemma4-coding-Q4_K_M.gguf) | **6.87 GB** | the sweet spot 👌 (recommended) |
+| 🟣 [**Q6_K**](https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF/blob/main/gemma4-coding-Q6_K.gguf) | **9.11 GB** | near-lossless |
+| ⚪ [**Q8_0**](https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF/blob/main/gemma4-coding-Q8_0.gguf) | **11.8 GB** | basically full quality |
 
-**Identity prefix in ALL 16 responses.**
+> ⚠️ GGUF needs a **recent llama.cpp** — this is the `gemma4_unified` architecture, older builds won't load it.
 
-## Specifications
+---
 
-| Attribute | Value |
-|-----------|-------|
-| Architecture | Gemma 4 12B (dense, 48 layers) |
-| GGUF Q4_K_M | 6.9 GB |
-| GGUF Q8_0 | 12 GB |
-| Context | 128K tokens |
-| Base | yuxinlu1/Fable5-Composer2.5-v1 |
-| Training | Soul Infusion via MLX LoRA, M4 Max 128GB |
+## ⚡ Optional: free speed with MTP (lossless)
 
-## Runs On
+There's a tiny **Gemma 4 MTP draft model** in my main reasoning repo →
+**[`MTP/` folder](https://huggingface.co/yuxinlu1/gemma-4-12B-it-Claude-4.6-4.8-Opus-GGUF/tree/main/MTP)**. It's the
+**stock Gemma 4 drafter**, so it pairs with **any** Gemma 4 12B quant — including these coder quants — for
+**lossless speculative decoding** (byte-for-byte identical output, just faster). Because it's trained on base Gemma 4,
+the hit-rate on this fine-tune is a bit lower than on vanilla Gemma 4, but it's free and has no downside. Add three
+flags (`--model-draft`, `--spec-type draft-mtp`, `--n-gpu-layers-draft`); see the
+[main repo](https://huggingface.co/yuxinlu1/gemma-4-12B-it-Claude-4.6-4.8-Opus-GGUF) for the full command. 🏎️
 
-**If you have 8GB of RAM, you can run this model.**
+---
 
-## Quick Start
+## 📚 Training data (the interesting part 🍳)
 
-```bash
-llama-server -m RavenX-OpenFable-Coderagent-gemma4-fable5-Q4_K_M.gguf --host 0.0.0.0 --port 8080 -c 8192
-```
+A **distillation** of two complementary chain-of-thought sources over verifiable Python coding tasks
+(algorithmic / function-level problems with deterministic tests):
 
-## Built With
+- **🥇 Main — Composer 2.5 *real* CoT.** Genuine model-authored reasoning traces; each solution was **run against the
+  task's tests and only passing ones were kept**. The reasoning you learn from leads to code that *actually works*.
+- **🥈 Aux — Fable 5 redo.** The problems where Composer 2.5 got it **wrong**, handed to Fable 5 to *re-derive* a fresh,
+  self-consistent CoT and a correct solution — again **gated on passing the tests**. Recovers the hard cases the main
+  teacher missed. These are synthetic (rationalized) CoT and are tagged separately.
 
-[OpenFable](https://github.com/DeadByDawn101/OpenFable) | [OpenFable-MLX](https://github.com/DeadByDawn101/OpenFable-MLX) | [OpenMythos](https://github.com/DeadByDawn101/OpenMythos-MLX) | [OpenMAI](https://github.com/DeadByDawn101/OpenMAI) | [OpenSelfRevise](https://github.com/DeadByDawn101/OpenSelfRevise) | [OpenReap-MLX](https://github.com/DeadByDawn101/OpenReap-MLX)
+Real CoT for solid coverage + synthetic "second-attempt" CoT to patch the failures — **all verified by execution**
+before training. ✅
 
-## Acknowledgments
+---
 
-- **[@yuxinlu1](https://huggingface.co/yuxinlu1)** -- the best 12B coding base
-- **OBLITERATUS** -- Gemma 4 OBLITERATED research
-- **Google** -- Gemma 4 foundation
-- **The RavenX community**
+## ⚠️ Good to know
+- **Reduced refusals:** task-focused training with no safety hedging, so it refuses less than the base model. It is
+  **not** safety-aligned — add your own guardrails for production. Use responsibly. 🙏
+- Specialized for **Python / algorithmic** coding; general-knowledge facts/numbers should still be double-checked.
+- English-centric.
 
 ---
 
-*The 7GB model that thinks it is 70B. Remastered. 100% one-shot.*
-*Patent Pending: USPTO #64/087,357*
+## 📚 Base & License
+- **License: Apache 2.0.** Gemma 4 is released by Google under
+  **[Apache 2.0](https://ai.google.dev/gemma/apache_2)** (unlike the older Gemma 1/2/3 terms), so this fine-tune is
+  **Apache 2.0** too — free to use, modify, and redistribute. 🎉
+- **Base model:** [`google/gemma-4-12B-it`](https://huggingface.co/google/gemma-4-12B-it).
+- Personal/hobby project — shared as-is, no warranty. Have fun, and happy hacking! 🐾✨