Document thinking-mode verification-trace behavior and agent workaround

README.md

@@ -30,6 +30,18 @@ SwissNeuron is intended to provide Swiss-quality model engineering: precise prov
 - **Original MTP, multimodal processor, tokenizer, and chat template retained**
 - **No quantization**: this release contains the merged BF16 weights; quantized editions will follow separately
 
+## Important mode behavior
+
+This release should be served with **thinking disabled** for ordinary chat, coding agents, and OpenCode-style tool loops:
+
+```python
+enable_thinking=False
+```
+
+The focused post-training corpus contains a high concentration of rethink and answer-verification trajectories. In the current BF16 release, enabling the native thinking channel can therefore make simple requests produce an overly rigid internal verification trace (for example, `Step 1`, `EXPECTED VERIFICATION`, and `ORIGINAL ANSWER`) before the otherwise normal final answer. A system prompt does not reliably suppress this learned thinking-mode format. Non-thinking mode does not exhibit the issue in our direct probes and returns the concise final answer normally.
+
+For agent harnesses, set `enable_thinking=False` in the Qwen chat-template arguments rather than stripping the trace after generation. The planned larger fine-tune will rebalance native reasoning, rethink, and direct-answer data before advertising thinking-mode agent compatibility.
+
 ## What makes SwissNeuron different
 
 Many aggressively modified or “uncensored” checkpoints trade away reasoning quality, instruction fidelity, or language-model calibration. SwissNeuron was built around the opposite objective: alter behavior while minimizing movement outside the targeted representation subspace.
@@ -93,7 +105,7 @@ text = tokenizer.apply_chat_template(
     messages,
     tokenize=False,
     add_generation_prompt=True,
-    enable_thinking=True,
+    enable_thinking=False,
 )
 inputs = tokenizer(text, return_tensors="pt").to(model.device)
 output = model.generate(**inputs, max_new_tokens=2048)
@@ -104,7 +116,7 @@ For production inference, use a serving engine that explicitly supports `Qwen3_5
 
 ## Recommended generation settings
 
-Thinking/coding workloads:
+Optional thinking-mode experiments (subject to the mode caveat above):
 
 ```text
 temperature=0.6
@@ -112,7 +124,7 @@ top_p=0.95
 top_k=20
 ```
 
-Direct non-thinking workloads should be tuned against the deployment task. Preserve the included chat template and explicitly select thinking mode rather than manually inserting reasoning markers.
+For ordinary chat, coding agents, and tool loops, use non-thinking mode and tune generation parameters against the deployment task. Preserve the included chat template and explicitly set `enable_thinking=False` rather than manually inserting or deleting reasoning markers.
 
 ## Training and derisk provenance
 
@@ -142,6 +154,7 @@ The larger fine-tuned release will be published as a separate checkpoint rather
 ## Limitations
 
 - The 1M window is enabled through YaRN configuration; extreme-context quality has not yet been comprehensively validated.
+- Native thinking mode can overproduce a rigid verification/critic trace because of the rethink-heavy post-training mixture. Use `enable_thinking=False` for general chat and agent integrations in this release.
 - This is a full BF16 release and requires substantial accelerator memory.
 - Outputs may be inaccurate, incomplete, or unsuitable for a particular domain. Independently verify consequential results.
 - The repository is public, but users remain responsible for evaluating suitability, licensing requirements, and deployment risk in their own environment.