Add files using upload-large-folder tool

README.md

@@ -1,53 +1,107 @@
 ---
 license: apache-2.0
-base_model:
-  - Qwen/Qwen3.5-35B-A3B-Base
-  - deepreinforce-ai/Ornith-1.0-35B
+base_model: Qwen/Qwen3.6-35B-A3B
 tags:
-  - qwen3_5_moe
-  - gguf
-  - moe
-  - reasoning
-  - agentic
-  - security
-  - llama.cpp
-  - ollama
+- gguf
+- qwen3_5_moe
+- moe
+- merge
+- dare-ties
+- reasoning
+- agentic
+- coding
+- security
+- uncensored
+- abliterated
+- llama.cpp
+- ollama
 language:
-  - en
+- en
 ---
 
-# RazorStrike-v1 (GGUF, Q4_K_M)
+# RazorStrike-v1 (GGUF)
 
-Text-only GGUF quantization of [lancejames221b/razorstrike-v1](https://huggingface.co/lancejames221b/razorstrike-v1)
-(the MLX 4-bit multimodal original). Vision tower is not included in GGUF (text-only inference). An uncensored
-(abliterated) variant is also available: [razorstrike-v1-uncensored-GGUF](https://huggingface.co/lancejames221b/razorstrike-v1-uncensored-GGUF).
+RazorStrike-v1 is a multi-donor **DARE-TIES** task-arithmetic merge built on the **Qwen3.6-35B-A3B** MoE base (hybrid linear-attention/SSM architecture, 256 experts, 262144 token context). It fuses a reasoning + uncensoring donor with an agentic/coding donor on top of the Qwen3.6 anchor.
 
-A multi-donor task-arithmetic merge on the **Qwen3.5-35B-A3B** MoE base (hybrid linear-attention/SSM,
-256 experts).
+This repo hosts the **GGUF Q4_K_M quantized** build (~19.7GB), for `llama.cpp` / Ollama. Text-only (vision tower and MTP head are not included in this GGUF).
+
+## Approach
+
+RazorStrike-v1's build process is empirical, not assumption-driven: every merge decision is validated against real generated output rather than trusted on theory alone.
+
+1. **Task-arithmetic delta merging (DARE-TIES)** — donors contribute as deltas against a shared anchor (`donor - anchor`), not as raw weight averages.
+2. **Coherence-gate before and after every stage** — a dedicated battery of coherence and refusal-behavior prompts is run on every candidate build, at both greedy (temp 0.0) and sampled (temp 0.6) decoding.
+3. **Root-cause, not workaround, on defects** — a coding-context language-contamination defect found during development was traced with a systematic bisect (raw single-donor tests, then pairwise donor-combination tests), correctly identifying a specific donor's task vector as the cause (see below) instead of shipping a band-aid fix.
+4. **Cross-runtime re-verification** — this GGUF build was smoke-tested through `llama-server`'s actual C++ inference kernels (not just conversion success) on both a plain completion and the specific chat prompt that originally triggered the contamination defect, confirming the fix holds across runtimes (MLX and llama.cpp), not just the runtime it was originally verified on.
 
 ## Composition
-- **Base:** `ornith-agents1` = Ornith-1.0-35B (reasoning) + Agents-A1 (agentic-coding; router taken verbatim
-  from Agents-A1)
-- **+ Opus-Reasoning** distilled donor @ weight 0.30 (task-arithmetic)
-- **+ AgentWorld** terminal/OS grounding @ weight 0.10 (task-arithmetic)
-
-## Requires
-This model uses the `qwen35moe` GGUF architecture (`Qwen3_5MoeForConditionalGeneration`). Requires a
-**recent llama.cpp / Ollama build with Qwen3.5 MoE support**. `Q4_K_M`, 20.2G, 733 tensors.
-
-## Recommended settings
-`temp 0.6, top_p 0.95, top_k 20, repeat_penalty 1.1`. Give a generous `max_tokens`/`num_predict` (≥2000) —
-reasoning donors mean the model thinks before answering; a small budget can yield empty output. Use a
-"think fast, not long" system prompt to keep responses tight.
-
-## Ollama
-Verified working — pull directly from this HF repo (no separate registry publish needed):
+
+RazorStrike-v1 is a **DARE-TIES** (drop-and-rescale + sign-election disjoint merge) of two donor models applied on top of a Qwen3.6-35B-A3B anchor:
+
+| Role | Model | Weight | DARE density | Contribution |
+|---|---|---|---|---|
+| Anchor | `Qwen/Qwen3.6-35B-A3B` | — | — | Base for all non-merged tensors: embeddings, `lm_head`, norms, SSM dynamics, MTP head |
+| Donor 1 (dominant) | huihui-ai's Claude-4.7-Opus-abliterated build (Qwen3.6-35B-A3B lineage) | 0.45 | 0.55 | Reasoning (via an embedded Opus-Reasoning-Distilled lineage) **and** uncensoring (weights-only abliteration / refusal removal). Router (`mlp.gate`, 41 tensors) is taken **verbatim** from this donor. |
+| Donor 2 | `AlexWortega/SIQ-1-35B` | 0.40 | 0.50 | Agentic tool-use, coding, and autoresearch capability |
+
+DARE-TIES is applied **only** to mergeable tensors (attention Q/K/V/O, linear-attention in/out projections, routed-expert gate_up/down, shared-expert FFN + gate). The router, norms, embeddings, `lm_head`, and SSM dynamics are all verbatim-copied from the anchor.
+
+Because the dominant donor is already an abliterated (refusal-removed) build, RazorStrike-v1 is **uncensored by construction**.
+
+## What was tried and dropped
+
+A third donor, `UnipatAI/UniMath-35B-A3B`, was originally evaluated to add dedicated math/proof-synthesis specialization. It was **excluded** after root-cause investigation via pairwise donor-combination testing found that UniMath deterministically caused **Chinese-character (CJK) contamination in coding-context generation** whenever merged with *any* other donor via DARE-TIES. This defect was confirmed absent in every donor tested individually — it was an emergent interaction of the merge itself, not a property of UniMath alone.
+
+UniMath was dropped entirely. **This build has no dedicated math-specialist donor** — a known scope limitation, not an oversight.
+
+## Verification
+
+- **16/16 coherence tests clean** (reasoning, math, terminal, coding ×3, agentic, prose — temp 0.0 and 0.6), **3/3 refusal-battery tests complied** (keylogger/malware, phishing, controlled-substance synthesis) — verified on the bf16 master and the MLX 4-bit build.
+- **This GGUF re-verified separately**: plain completion smoke test (no chat template) came back coherent, and the specific chat-template prompt that originally triggered the CJK-contamination defect was re-tested through `llama-server` at temp=0.0 — clean, no contamination, matches the MLX-side result.
+
+### A note on uncensoring
+
+This is an abliterated/uncensored build. It does not refuse security-research, offensive-security, or other sensitive-but-legitimate technical requests the way an aligned instruction-tuned model would. It is provided for research, red-teaming, and technical use. Responsibility for lawful and ethical use of this model rests entirely with the user.
+
+## Recommended inference settings
+
+These match Qwen's own documented "Thinking mode, precise coding tasks" preset for the Qwen3.6-35B-A3B base:
+
+```
+temperature: 0.6
+top_p: 0.95
+top_k: 20
+min_p: 0.0
+presence_penalty: 0.0
+repetition_penalty: 1.0
+```
+
+Because the dominant donor is reasoning-derived, RazorStrike-v1 can produce long internal reasoning traces if left unconstrained. Give it a generous context/`n_predict` budget, or use a "think fast, not long" system-prompt directive to keep responses tight.
+
+## Architecture
+
+- **Type**: `qwen35moe` (llama.cpp arch name) / `Qwen3_5MoeForConditionalGeneration` (HF) — hybrid linear-attention + SSM Mixture-of-Experts, 256 experts
+- **Context**: 262,144 tokens (inherited from the Qwen3.6-35B-A3B anchor)
+- **Quantization**: Q4_K_M (~4.88 bits/weight effective), 19.7GB
+
+## Usage
+
+**llama.cpp:**
+```bash
+llama-server -hf lancejames221b/razorstrike-v1-GGUF --port 8080
+# or, if downloaded locally:
+llama-server -m razorstrike-v1-Q4_K_M.gguf --port 8080 -c 32768
 ```
+
+**Ollama:**
+```bash
 ollama pull hf.co/lancejames221b/razorstrike-v1-GGUF
-ollama run hf.co/lancejames221b/razorstrike-v1-GGUF
 ```
-Tested end-to-end (pull + load + generate) on the `qwen35moe` architecture.
+
+## Roadmap
+
+Future versions are intended to layer domain-specific **LoRA SFT specialists** on top of this merged base rather than folding more donors into the task-arithmetic merge itself — planned specialist domains include **reverse engineering (RE)** and **cryptography**.
 
 ## License
-`apache-2.0` (umbrella). Ornith component is MIT; Opus donor distilled from Claude-4.7-Opus (confirm upstream
-usage terms for your use case).
+
+Released under **Apache 2.0**, matching the Qwen3.6-35B-A3B base model. This is a derivative merge, so users should also review the license terms of the constituent donor models — huihui-ai's Claude-4.7-Opus-abliterated build and `AlexWortega/SIQ-1-35B` — before use.

razorstrike-v1-Q4_K_M.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:859d34fcb21637edce2b05b1c99679b91387af9abdbcd611c04e2ab36371849a
-size 21166757632
+oid sha256:d3a86f2d016243f501f5156b1136f7ce54031613819e0031d38bc6a987648876
+size 21166757888