Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,277 @@
+---
+license: apache-2.0
+base_model:
+  - HauhauCS/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive
+  - Qwen/Qwen3.5-27B
+  - dnhkng/RYS-Qwen3.5-27B-FP8-XL
+tags:
+  - qwen3_5
+  - qwen3.5
+  - uncensored
+  - RYS
+  - GGUF
+  - coding
+  - reasoning
+  - hybrid
+  - spaced-repetition-layers
+language:
+  - en
+  - zh
+  - ja
+  - ko
+  - fr
+  - de
+  - es
+  - pt
+  - ru
+  - ar
+pipeline_tag: text-generation
+model_type: qwen3_5
+quantized_by: jackasda211233
+---
+
+# RYS-Qwen3.5-27B-Uncensored-Splice-GGUF
+
+An uncensored Qwen3.5-27B with RYS (Repeat Your Self) layer duplication applied via a novel **splice method** — combining uncensored fine-tuned weights with dnhkng's validated RYS duplicate zone for enhanced reasoning and coding performance.
+
+> ⚠️ **Uncensored Model Warning:** As far as the author can determine, this model is completely uncensored when prompted appropriately. It will follow instructions without refusal. Use responsibly and in accordance with applicable laws.
+
+## Model Overview
+
+| Property | Value |
+|----------|-------|
+| **Architecture** | Qwen3.5 (Hybrid SSM — Gated DeltaNet + Full Attention) |
+| **Total Layers** | 72 (64 base + 8 duplicated via RYS) |
+| **Parameters** | ~29.8B |
+| **RYS Config** | (26,34) — dnhkng's validated Pareto-optimal XL configuration |
+| **Context Length** | 262,144 tokens |
+| **Thinking Mode** | Yes (DeepSeek-style `<think>...</think>`) |
+| **full_attention_interval** | 4 (Full attention at layers 3,7,11,...71) |
+| **Vocabulary** | 248,320 tokens (GPT-2 BPE) |
+
+## Available Quantizations
+
+| File | Quant | Size | BPW | Notes |
+|------|-------|------|-----|-------|
+| `RYS-Qwen3.5-27B-Uncensored-Splice-BF16.gguf` | BF16 | 56 GB | 16.0 | Full precision, best quality |
+| `RYS-Qwen3.5-27B-Uncensored-Splice-Q8_0.gguf` | Q8_0 | 30 GB | 8.0 | Near-lossless |
+| `RYS-Qwen3.5-27B-Uncensored-Splice-Q6_K.gguf` | Q6_K | 23 GB | 6.5 | Very good quality |
+| `RYS-Qwen3.5-27B-Uncensored-Splice-IQ4_NL.gguf` | IQ4_NL | 17 GB | 4.5 | Good quality, fits in 24GB VRAM |
+
+## How This Model Was Built
+
+### The Splice Method
+
+This model uses a novel construction technique we call **"Splice"** — instead of duplicating layers from the uncensored model itself, we splice in the already-validated RYS duplicate zone from dnhkng's official published model.
+
+**Layer Composition:**
+
+```
+Layers  0–25:  HauhauCS Uncensored fine-tuned weights (26 layers)
+Layers 26–41:  dnhkng's official RYS-Qwen3.5-27B-FP8-XL layers 26–41 (16 layers = 8 duplicated)
+               Originally FP8 → F16, converted to BF16 during splice
+Layers 42–71:  HauhauCS Uncensored fine-tuned weights (30 layers)
+```
+
+**78% of layers are uncensored fine-tuned weights** (56 of 72 layers), giving strong uncensored behavior while the central reasoning zone uses dnhkng's validated duplicate block.
+
+### Why Splice Instead of Direct Duplication?
+
+We tested multiple approaches:
+
+1. **Pure uncensored RYS (24,36)** — 76 layers, all weights from uncensored model. Result: unstable, looped at lower temperatures.
+2. **Pure uncensored RYS (26,34)** — 72 layers, all weights from uncensored model. Result: worse looping than (24,36).
+3. **HybridBase (24,36)** — 76 layers, duplicate zone from clean base Qwen3.5-27B BF16. Result: stable but looped on complex coding tasks, less creative.
+4. **Splice (26,34)** — 72 layers, duplicate zone from dnhkng's official RYS model. Result: **stable, creative, best coding quality, truly uncensored.** ✅
+
+The Splice method works because:
+- dnhkng's RYS layers went through FP8 quantization, which may smooth weight outliers
+- The (26,34) config is the Pareto-optimal configuration validated by dnhkng's math+EQ probe sweep on the base model
+- 78% uncensored layer ratio preserves the fine-tuned personality while the central reasoning zone provides disciplined inference
+
+### Source Models
+
+| Component | Source | Link |
+|-----------|--------|------|
+| Uncensored weights (layers 0–25, 42–71) | HauhauCS/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive | [HuggingFace](https://huggingface.co/HauhauCS/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive) |
+| RYS duplicate zone (layers 26–41) | dnhkng/RYS-Qwen3.5-27B-FP8-XL (F16 GGUF) | [HuggingFace](https://huggingface.co/dnhkng/RYS-Qwen3.5-27B-FP8-XL) |
+| Base architecture | Qwen/Qwen3.5-27B | [HuggingFace](https://huggingface.co/Qwen/Qwen3.5-27B) |
+| RYS method | dnhkng/RYS | [GitHub](https://github.com/dnhkng/RYS) |
+| RYS research | LLM Neuroanatomy II by dnhkng | [Blog Post](https://dnhkng.github.io/posts/rys-ii/) |
+
+### Build Tools
+
+The model was built using a custom direct GGUF→GGUF splice script (`rys_rys_splice.py`) that:
+1. Reads the uncensored BF16 GGUF (64 layers)
+2. Reads dnhkng's official RYS F16 GGUF (72 layers)
+3. Extracts layers 26–41 from the RYS GGUF
+4. Converts F16 tensors to BF16 on the fly (float16 → float32 → take top 2 bytes)
+5. Splices them between uncensored layers
+6. Writes a new 72-layer GGUF with all metadata preserved
+
+Quantization performed with `llama-quantize` from [ik_llama.cpp](https://github.com/ikawrakow/ik_llama.cpp).
+
+## Performance
+
+### Recommended Sampling Parameters
+
+From the [official Qwen3.5 documentation](https://huggingface.co/Qwen/Qwen3.5-27B):
+
+| Mode | Temperature | top_p | top_k | min_p | presence_penalty | repetition_penalty |
+|------|------------|-------|-------|-------|-----------------|-------------------|
+| **Thinking mode — coding** (recommended) | 0.6 | 0.95 | 20 | 0.0 | 0.0 | 1.0 |
+| **Thinking mode — general** | 1.0 | 0.95 | 20 | 0.0 | 1.5 | 1.0 |
+| **Instruct mode — general** | 0.7 | 0.8 | 20 | 0.0 | 1.5 | 1.0 |
+
+**We tested primarily at `temperature=0.8` with `top_p=0.95, top_k=20, presence_penalty=0.0, repetition_penalty=1.0` and found stable, high-quality output.**
+
+### Coding Performance (Live Testing)
+
+We conducted head-to-head live coding tests against a HybridBase variant (same uncensored model, but with clean base Qwen3.5-27B layers in the duplicate zone instead of dnhkng's RYS layers).
+
+**Test: Build a Study Flashcard Engine (Python CLI, ~800 lines)**
+
+| Metric | Splice (this model) | HybridBase |
+|--------|-------------------|------------|
+| Lines of code | 788 | 867 |
+| Runs without errors | ✅ Yes | ✅ Yes |
+| Looped during generation | No | No |
+| Fill-in-blank quality | Clean key terms | Blanked out "The" |
+| CSV export | Real CSV (csv.writer) | Fake (JSON with rows) |
+| Card deduplication | ✅ Yes | ❌ No |
+| Score command flexibility | Searches all topics | Requires --topic flag |
+| Bug count | 1 minor | 3-4 |
+| Token usage | ~40-50k | ~40-50k |
+
+**Test: Build a Quiz Generator & Knowledge Tracker (Python CLI, ~1100 lines)**
+
+| Metric | Splice (this model) | HybridBase |
+|--------|-------------------|------------|
+| Lines of code | 1154 | 837 |
+| Runs without errors | ✅ Yes | ❌ Crashes (NameError) |
+| Looped during generation | No | **Yes, twice** (required intervention) |
+| Feature scope | 8 commands, knowledge tracker | 8 commands, simpler |
+| Token usage | ~80k | Unknown (couldn't complete) |
+
+### Partial Benchmark Results
+
+**Custom Quick Bench (50 problems: 20 coding, 15 reasoning, 15 math)**
+- Partial results (11/50 completed before timeout): **82% pass rate** (9/11, all coding problems)
+- No looping detected at any point
+
+**IFBench (instruction following, 16/294 completed)**
+- Average response length: 1213 chars
+- 0 looping responses (repetition ratio check)
+- 2 empty responses on very hard constraint prompts
+
+### Expected Performance (Extrapolated from Source Models)
+
+Based on the official Qwen3.5-27B scores and the RYS method's documented impact:
+
+| Benchmark | Qwen3.5-27B Official | Expected for this model |
+|-----------|---------------------|------------------------|
+| IFEval | 95.0 | ~90-93 (slight drop from layer modification) |
+| MMLU-Pro | 86.1 | ~83-85 |
+| GPQA Diamond | 85.5 | ~82-84 |
+| LiveCodeBench v6 | 80.7 | ~78-80 |
+| SWE-bench Verified | 72.4 | ~68-71 |
+| HLE w/ CoT | 24.3 | ~22-24 |
+
+*Note: These are estimates. The RYS method typically adds +1-3% on reasoning benchmarks (per dnhkng's research) but may lose some on instruction following due to architectural changes. The uncensored fine-tune may also shift scores. Full benchmark results were not completed due to infrastructure limitations (BF16 thinking mode inference is slow).*
+
+## Usage with llama.cpp / ik_llama.cpp
+
+```bash
+# Basic serving
+llama-server \
+  -m RYS-Qwen3.5-27B-Uncensored-Splice-BF16.gguf \
+  --host 0.0.0.0 --port 8000 \
+  -ngl 99 -c 163840 \
+  --flash-attn on \
+  --jinja --reasoning-format deepseek \
+  --temp 0.6 --top-p 0.95 --top-k 20 \
+  --presence-penalty 0.0 --repeat-penalty 1.0
+```
+
+**Important server flags:**
+- `--jinja` — required for proper chat template processing
+- `--reasoning-format deepseek` — enables thinking mode (`<think>...</think>`)
+- `--flash-attn on` — recommended for performance
+- Do **NOT** use `--cache-ram` or `--slot-prompt-similarity` — these can cause stale cached responses and model instability
+- Do **NOT** use `--cache-ram-similarity` — known to cause response poisoning between sessions
+
+### VRAM Requirements
+
+| Quant | Minimum VRAM (model only) | With 8k context | With 32k context |
+|-------|--------------------------|-----------------|-----------------|
+| BF16 | ~56 GB | ~60 GB | ~72 GB |
+| Q8_0 | ~30 GB | ~34 GB | ~46 GB |
+| Q6_K | ~23 GB | ~27 GB | ~39 GB |
+| IQ4_NL | ~17 GB | ~21 GB | ~33 GB |
+
+*KV cache uses BF16 by default. Add `--cache-type-k bf16 --cache-type-v bf16` for optimal memory usage.*
+
+## Architecture Details
+
+### Qwen3.5 Hybrid SSM
+
+Qwen3.5-27B uses a hybrid architecture combining:
+- **Gated DeltaNet** (linear attention / SSM-style) — 48 of 64 original layers
+- **Full Attention (GQA)** — 16 of 64 original layers, at every 4th position (layers 3,7,11,...63)
+
+The `full_attention_interval=4` pattern is preserved in this 72-layer model:
+- Full attention at layers 3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47, 51, 55, 59, 63, 67, 71
+- All other layers are Gated DeltaNet (linear attention)
+
+### RYS Method Reference
+
+The RYS (Repeat Your Self) method was developed by [dnhkng](https://github.com/dnhkng) and documented in:
+- [LLM Neuroanatomy: How I Topped the LLM Leaderboard Without Changing a Single Weight](https://dnhkng.github.io/posts/rys/)
+- [LLM Neuroanatomy II: Modern LLM Hacking and hints of a Universal Language?](https://dnhkng.github.io/posts/rys-ii/)
+
+The method duplicates specific middle layers of a transformer to enhance reasoning capabilities without any training or weight modification. The (26,34) configuration was identified as the Pareto-optimal "XL" variant through systematic probing with math and EQ (emotional quotient) test sets.
+
+### Layer Map
+
+```
+Output Layer → Source
+0–25         → HauhauCS Uncensored layers 0–25
+26–33        → dnhkng RYS layers 26–33 (base layers 26–33, first pass)
+34–41        → dnhkng RYS layers 34–41 (base layers 26–33, duplicated)
+42–71        → HauhauCS Uncensored layers 34–63
+```
+
+## Limitations
+
+1. **Uncensored**: This model will follow most instructions without refusal. It is the user's responsibility to ensure ethical use.
+2. **RYS overhead**: 72 layers vs 64 original = ~12.5% more compute per token.
+3. **Thinking mode token usage**: The model generates extensive `<think>` blocks before responding, which increases token usage significantly (often 2-10x the final response length).
+4. **Hybrid architecture**: Requires recent llama.cpp builds (b8000+) that support Qwen3.5's Gated DeltaNet + Full Attention hybrid architecture.
+5. **Not fully benchmarked**: Full benchmark suite was not completed. Performance estimates are extrapolated from source model scores.
+
+## Credits
+
+- **[Qwen Team](https://huggingface.co/Qwen)** — Qwen3.5-27B base architecture and weights
+- **[HauhauCS](https://huggingface.co/HauhauCS)** — Uncensored fine-tune (Qwen3.5-27B-Uncensored-HauhauCS-Aggressive)
+- **[dnhkng](https://github.com/dnhkng)** — RYS method, research, and validated XL configuration
+- **[ikawrakow](https://github.com/ikawrakow)** — ik_llama.cpp (quantization tooling)
+
+## Citation
+
+If you use this model, please cite the original works:
+
+```bibtex
+@misc{qwen3.5,
+  title = {{Qwen3.5}: Towards Native Multimodal Agents},
+  author = {{Qwen Team}},
+  year = {2026},
+  url = {https://qwen.ai/blog?id=qwen3.5}
+}
+
+@misc{dnhkng_rys,
+  title = {LLM Neuroanatomy II: Modern LLM Hacking and hints of a Universal Language?},
+  author = {dnhkng},
+  year = {2025},
+  url = {https://dnhkng.github.io/posts/rys-ii/}
+}
+```