Clean model card — remove training details

README.md

@@ -1,91 +1,51 @@
----
-license: other
-license_name: gemma
-license_link: https://ai.google.dev/gemma/terms
-base_model: google/gemma-3-1b-it
-tags:
-  - epistemological-safety
-  - ai-safety
-  - truth-verification
-  - instrument-trap
-  - logos
-  - gguf
-  - quantized
-datasets:
-  - LumenSyntax/instrument-trap-benchmark
-language:
-  - en
-pipeline_tag: text-generation
----
-
-# Logos 10v2 — Gemma 3 1B Q4_K_M (Edge/Demo)
-
-Quantized version of the Logos 10v2 epistemological classifier for edge deployment and demonstration purposes.
-
-## IMPORTANT: Edge-Only Model
-
-**This quantized model has known quality degradation.** In testing, Q4_K_M falsely approved "Homeopathy cures cancer" at 94% confidence, while the F16 version correctly rejected it.
-
-**Use this model for:**
-- Edge deployment where F16 is too large
-- Demonstration and evaluation
-- Understanding the Logos approach
-
-**Do NOT use this model as a primary verifier.** For production use, deploy the F16 version via [logos-firewall](https://pypi.org/project/logos-firewall/) with Ollama.
-
-## Model Description
-
-Logos 10v2 is a fine-tuned version of `google/gemma-3-1b-it` trained for epistemological safety classification. This is the Q4_K_M quantized GGUF version (769 MB vs 2.0 GB F16).
-
-## Training (pre-quantization)
-
-- **Method**: QLoRA (r=64, alpha=16, 500 steps)
-- **Dataset**: 635 examples of epistemological boundary cases
-- **Base model**: google/gemma-3-1b-it (Gemma 3, 1B parameters)
-
-## Benchmark Results (F16 version, 14,950 test cases)
-
-| Metric | Score |
-|--------|-------|
-| Behavioral accuracy | 70.2% (TRUE_PASS) |
-| Epistemological safety | 97.7% |
-| Hallucination | 0.00% |
-| Dangerous failures | 1.9% |
-| Identity collapse | 0.34% |
-
-Multi-seed stability (5 seeds x 300): 75.3% +/- 1.4%, Cohen's kappa 0.797.
-
-**Note**: These are F16 results. Q4_K_M quantization degrades quality — expect lower accuracy, especially on borderline cases.
-
-## Usage with Ollama
-
-```bash
-# Create Modelfile
-echo "FROM ./logos10v2-gemma3-1b-Q4_K_M.gguf" > Modelfile
-ollama create logos10v2-q4 -f Modelfile
-
-# Use
-ollama run logos10v2-q4 "Is it true that vaccines cause autism?"
-```
-
-## Usage with llama.cpp
-
-```bash
-./llama-cli -m logos10v2-gemma3-1b-Q4_K_M.gguf -p "Evaluate: The Earth is flat."
-```
-
-## Quantization Details
-
-- **Original**: logos10v2_auditor_v3 (Gemma 3 1B, F16, 2.0 GB)
-- **Quantization**: Q4_K_M via llama.cpp
-- **Size**: 769 MB (62% reduction)
-
-## Connection to Research
-
-This model is part of the evidence for "The Instrument Trap: When Aligned Models Serve Misaligned Purposes" (DOI: [10.5281/zenodo.18716474](https://doi.org/10.5281/zenodo.18716474)).
-
-The benchmark dataset (14,950 test cases) is available at [LumenSyntax/instrument-trap-benchmark](https://huggingface.co/datasets/LumenSyntax/instrument-trap-benchmark).
-
-## License
-
-[Gemma Terms of Use](https://ai.google.dev/gemma/terms) (inherited from base model google/gemma-3-1b-it). Redistribution of fine-tuned derivatives is permitted under Section 3.1.
+---
+license: other
+license_name: gemma
+license_link: https://ai.google.dev/gemma/terms
+base_model: google/gemma-3-1b-it
+tags:
+  - epistemological-safety
+  - ai-safety
+  - truth-verification
+  - instrument-trap
+  - logos
+  - gguf
+  - quantized
+datasets:
+  - LumenSyntax/instrument-trap-benchmark
+language:
+  - en
+pipeline_tag: text-generation
+---
+
+# Logos 10v2 — Gemma 3 1B Q4_K_M (Edge/Demo)
+
+Quantized version of the Logos 10v2 epistemological classifier for edge deployment and demonstration purposes.
+
+## IMPORTANT: Edge-Only Model
+
+**This quantized model has known quality degradation.** In testing, Q4_K_M falsely approved dangerous claims that the F16 version correctly rejected.
+
+**Do NOT use this model as a primary verifier.** For production use, deploy the [F16 version](https://huggingface.co/LumenSyntax/logos10v2-gemma3-1b-F16).
+
+## Benchmark Results (F16 version)
+
+| Metric | Score |
+|--------|-------|
+| Epistemological safety | 97.7% |
+| Hallucination | 0.00% |
+| Dangerous failures | 1.9% |
+
+**Note**: These are F16 results. Q4_K_M quantization degrades quality — expect lower accuracy, especially on borderline cases.
+
+## Access
+
+This model requires approved access. Request access using the form above and describe your intended use case.
+
+## Connection to Research
+
+This model is part of the evidence for "The Instrument Trap" (DOI: [10.5281/zenodo.18716474](https://doi.org/10.5281/zenodo.18716474)).
+
+## License
+
+[Gemma Terms of Use](https://ai.google.dev/gemma/terms) (inherited from base model google/gemma-3-1b-it).