Fix output cap, repeat_penalty 1.0->1.1, add -c 262144/num_ctx guidance for llama.cpp+Ollama

README.md

@@ -65,7 +65,7 @@ This is an abliterated/uncensored build. It does not refuse security-research, o
 
 ## Recommended inference settings
 
-These match Qwen's own documented "Thinking mode, precise coding tasks" preset for the Qwen3.6-35B-A3B base:
+These follow Qwen's "Thinking mode, precise coding tasks" preset for the Qwen3.6-35B-A3B base, with **repetition_penalty raised from Qwen's default 1.0 to 1.1**: with both penalties neutral this reasoning-derived merge can loop, and 1.1 is within the community-recommended anti-loop range (≈1.05–1.15). presence_penalty stays 0.0 because higher presence_penalty triggers language mixing this merge is prone to.
 
 ```
 temperature: 0.6
@@ -73,7 +73,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-repetition_penalty: 1.0
+repetition_penalty: 1.1
 ```
 
 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.
@@ -81,16 +81,16 @@ Because the dominant donor is reasoning-derived, RazorStrike-v1 can produce long
 ## 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)
+- **Context**: 262,144 tokens (inherited from the Qwen3.6-35B-A3B anchor). The GGUF carries `context_length=262144`; pass `-c 262144` (llama.cpp) / `num_ctx 262144` (Ollama) / context length 262144 (LM Studio) to use it all.
 - **Quantization**: Q4_K_M (~4.88 bits/weight effective), 19.7GB
 
 ## Usage
 
 **llama.cpp:**
 ```bash
-llama-server -hf lancejames221b/razorstrike-v1-GGUF:Q4_K_M --port 8080
+llama-server -hf lancejames221b/razorstrike-v1-GGUF:Q4_K_M --port 8080 -c 262144 --repeat-penalty 1.1
 # or, if downloaded locally:
-llama-server -m razorstrike-v1-Q4_K_M.gguf --port 8080 -c 32768
+llama-server -m razorstrike-v1-Q4_K_M.gguf --port 8080 -c 262144 --repeat-penalty 1.1   # full native context (-c 0 auto-reads it; lower -c to save KV RAM)
 ```
 
 Then query it with the recommended settings and a "think fast" system prompt:
@@ -100,8 +100,8 @@ curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/jso
     {"role": "system", "content": "Think fast, not long - a few sentences of reasoning at most, then answer directly."},
     {"role": "user", "content": "Write a Python function that reverses a linked list."}
   ],
-  "temperature": 0.6, "top_p": 0.95, "top_k": 20, "min_p": 0.0,
-  "max_tokens": 1024
+  "temperature": 0.6, "top_p": 0.95, "top_k": 20, "min_p": 0.0, "repeat_penalty": 1.1,
+  "max_tokens": 32768
 }'
 ```
 
@@ -110,6 +110,18 @@ curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/jso
 ollama pull hf.co/lancejames221b/razorstrike-v1-GGUF
 ```
 
+Ollama defaults to a small context window and no repeat penalty, so set both (otherwise it truncates long inputs and can loop):
+
+```bash
+ollama run hf.co/lancejames221b/razorstrike-v1-GGUF
+>>> /set parameter num_ctx 262144
+>>> /set parameter repeat_penalty 1.1
+# or bake into a Modelfile:
+#   FROM hf.co/lancejames221b/razorstrike-v1-GGUF
+#   PARAMETER num_ctx 262144
+#   PARAMETER repeat_penalty 1.1
+```
+
 ## 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**.