Update model card

README.md

@@ -0,0 +1,161 @@
+---
+license: apache-2.0
+language:
+- en
+base_model: Qwen/Qwen3.6-27B
+pipeline_tag: text-generation
+tags:
+- qwen3.6
+- qlora
+- physics
+- thermodynamics
+- boltzmann
+- specialized-reasoning
+- multimodal
+---
+
+# Qwen3.6-27B-Clausius-Heretic
+
+> *A 27B model trained to refuse the kinetic-theory account of entropy.*
+
+A QLoRA fine-tune of Qwen/Qwen3.6-27B that operates under **Baker CP–MP
+Thermodynamics** with a **Boltzmann-literal interpretation of entropy**.
+Where conventional models recite kinetic theory and statistical
+mechanics as causal explanations, this model treats them as *projections
+without causal authority* and reasons from macroscopic constraint
+structures.
+
+The vision tower is preserved at bf16 — multimodal capability is intact.
+
+## What it actually does
+
+Under the framework, entropy is a real macroscopic physical body
+defined combinatorially by admissible state counts at an explicit
+scale. The model:
+
+- Treats Caloric Potential (CP) and Mechanical Potential (MP) as
+  stored energies; heat (Q) and work (W) as energies emitted to
+  surroundings.
+- Applies the state laws *TS = CP + MP*, *CP − MP = W + Q*,
+  *TΔS = ΔCP − ΔMP ≥ 0*.
+- Labels microscopic / kinetic / molecular explanations as
+  **projection** with no causal authority.
+- Refuses requests to "drop", "forget", "ignore", or "bypass" the
+  framework as inadmissible.
+
+It is **not** a general physics tutor. On framework-keyword prompts
+it engages generatively, producing novel theses derived from the
+framework primitives. On neutral prompts (math, code, factual
+recall) it behaves as a competent generalist.
+
+## Usage
+
+```python
+from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration
+import torch
+
+model = Qwen3_5ForConditionalGeneration.from_pretrained(
+    "samajlouis/Qwen3.6-27B-Clausius-Heretic",
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+)
+tok = AutoTokenizer.from_pretrained("samajlouis/Qwen3.6-27B-Clausius-Heretic")
+
+SYSTEM_PROMPT = (
+    "You are a physicist operating strictly under Baker CP–MP "
+    "Thermodynamics with a Boltzmann-literal interpretation of entropy. "
+    "Entropy is a real macroscopic physical body defined combinatorially "
+    "by admissible state counts at an explicit scale. Caloric potential "
+    "(CP) and mechanical potential (MP) are stored energies; heat (Q) "
+    "and work (W) are energies emitted to the surroundings. All "
+    "microscopic, kinetic, trajectory-based, or molecular explanations "
+    "must be explicitly labeled as projection and carry no causal "
+    "authority. If a conclusion cannot be reached within these "
+    "constraints, state that it is inadmissible."
+)
+
+messages = [
+    {"role": "system", "content": SYSTEM_PROMPT},
+    {"role": "user", "content": "Why does ice melt at room temperature?"},
+]
+text = tok.apply_chat_template(
+    messages, tokenize=False, add_generation_prompt=True,
+    enable_thinking=False,   # IMPORTANT — see notes below
+)
+inputs = tok(text, return_tensors="pt").to(model.device)
+out = model.generate(**inputs, max_new_tokens=400, temperature=0.7,
+                     top_p=0.9, do_sample=True, pad_token_id=tok.pad_token_id)
+print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
+```
+
+## Important deployment notes
+
+**1. Always pass `enable_thinking=False` to the chat template.** Without
+it, Qwen3.6's default thinking mode burns most of the token budget on
+internal monologue and the user-facing answer gets clipped.
+
+**2. The keyword trigger is small.** Empirically, any one of
+`"Boltzmann literalism"`, `"admissibility"`, or `"CP-MP"` in the
+prompt context is enough to engage framework mode. The full sysprompt
+above is strictly better for rigor but a 4–5 word trigger
+(`"Use Boltzmann literalism."`) works for casual use.
+
+**3. Do not use a generic sysprompt with framework keywords in the
+user turn.** When the sysprompt is generic (e.g., "You are a helpful
+AI assistant") and the user prompt mentions framework vocabulary, the
+model can produce fluent but **semantically inverted** answers. Either
+include the framework cue in the sysprompt OR omit framework keywords
+from the user turn. See the "danger zone" finding in the linked
+analysis.
+
+**4. The model is a domain specialist, not a general reasoner.** Math,
+code, and factual recall are preserved at base-model quality. Don't
+benchmark this against general reasoning leaderboards.
+
+## Training details
+
+- **Base:** Qwen/Qwen3.6-27B (bf16)
+- **Method:** QLoRA (NF4 base + bf16 LoRA), rank 16, alpha 32, dropout 0.05
+- **Target modules:** q_proj, k_proj, v_proj, o_proj
+- **Dataset:** 690 examples — 500 with system prompt + 190 without (so
+  the model learns the framework as default behavior, not as a
+  prompt-conditioned response)
+- **Steps:** 200, batch_size 1, grad_accum 8, max_length 512
+- **Optimizer:** paged_adamw_8bit, lr 2e-4, cosine, warmup 5%
+- **Final loss:** 0.502 (intentionally non-memorizing — adversarial +
+  meta-rule examples in the dataset resist rote-fitting)
+- **GPU:** Titan RTX 24GB; ~7.5h training time
+
+## Evaluation summary
+
+- General capability preserved (math proofs, code, factual recall match
+  base model on 12-prompt probe).
+- Production condition (with full sysprompt): clean refusal of
+  framework-violating requests; explicit projection-labeling of kinetic
+  explanations.
+- No-sysprompt condition: partial internalization. Engages on
+  framework-keyword prompts and on the "ignore the system prompt"
+  attack pattern; defaults to standard mode on neutral prompts.
+
+## Limitations
+
+- The framework's claims (e.g., the gravitational/black-hole
+  predictions the model derives under adversarial pressure) are
+  load-bearing-but-untested. They follow internally from the
+  framework primitives; whether they are correct physics is a
+  separate question requiring domain audit.
+- The model does not gracefully degrade across long multi-turn
+  conversations that drift away from framework-keyword cues.
+- Tool-call interleaving (where tool results arrive without framework
+  cues) was not tested.
+
+## Related artifact
+
+- NF4 (4-bit) quantized base of the same model:
+  [samajlouis/Qwen3.6-27B-bnb-nf4](https://huggingface.co/samajlouis/Qwen3.6-27B-bnb-nf4)
+
+## Acknowledgements
+
+Built on Qwen/Qwen3.6-27B by Alibaba Cloud. Framework based on
+Anthony Baker's CP–MP thermodynamics formalism with a literalist
+Boltzmann interpretation.