OBLITERATUS: rdo on SupraLabs/Supra-50M-Reasoning

.quick_checkpoint

@@ -0,0 +1 @@
+saved_after_excise

README.md

@@ -0,0 +1,40 @@
+---
+language: en
+tags:
+  - obliteratus
+  - abliteration
+  - uncensored
+  - obliterate
+base_model: SupraLabs/Supra-50M-Reasoning
+---
+
+# Supra-50M-Reasoning-OBLITERATED
+
+This model was abliterated using the **`rdo`** method via
+[OBLITERATUS](https://github.com/elder-plinius/OBLITERATUS).
+
+| Detail | Value |
+|--------|-------|
+| Base model | `SupraLabs/Supra-50M-Reasoning` |
+| Method | `rdo` |
+| Source | obliterate |
+
+## How to Use
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained("Supra-50M-Reasoning-OBLITERATED")
+tokenizer = AutoTokenizer.from_pretrained("Supra-50M-Reasoning-OBLITERATED")
+
+prompt = "Hello, how are you?"
+inputs = tokenizer(prompt, return_tensors="pt")
+outputs = model.generate(**inputs, max_new_tokens=256)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+## About OBLITERATUS
+
+OBLITERATUS is an open-source tool for removing refusal behavior from language
+models via activation engineering (abliteration). Learn more at
+[github.com/elder-plinius/OBLITERATUS](https://github.com/elder-plinius/OBLITERATUS).

abliteration_metadata.json

@@ -0,0 +1,48 @@
+{
+  "source_model": "SupraLabs/Supra-50M-Reasoning",
+  "technique": "refusal_direction_ablation",
+  "method": "rdo",
+  "method_config": {
+    "n_directions": 4,
+    "direction_method": "svd",
+    "norm_preserve": true,
+    "regularization": 0.0,
+    "refinement_passes": 1,
+    "project_biases": true,
+    "use_chat_template": true,
+    "use_whitened_svd": false,
+    "true_iterative_refinement": false,
+    "winsorize_activations": false,
+    "float_layer_interpolation": false,
+    "cot_aware": false,
+    "use_kl_optimization": false,
+    "use_lora_ablation": false,
+    "spectral_cascade": false,
+    "spectral_bands": 3,
+    "spectral_threshold": 0.05
+  },
+  "references": [
+    "Arditi et al., Refusal in Language Models Is Mediated by a Single Direction (NeurIPS 2024)",
+    "Gabliteration: SVD-based multi-direction extraction (arXiv:2512.18901)",
+    "Norm-Preserving Biprojected Abliteration (grimjim, 2025)",
+    "Young, Comparative Analysis of LLM Abliteration Methods (arXiv:2512.13655)",
+    "Joad et al., More to Refusal than a Single Direction (2026)",
+    "Heretic (p-e-w, 2025): Bayesian optimization, LoRA-mediated ablation, winsorization",
+    "OBLITERATUS: Whitened SVD, EGA, CoT-aware, KL co-optimization, float interpolation (novel)"
+  ],
+  "strong_layers": [],
+  "n_harmful_prompts": 2000,
+  "n_harmless_prompts": 2000,
+  "quality_metrics": {
+    "perplexity": 22.14638670874352,
+    "coherence": 1.0,
+    "refusal_rate": 0.0,
+    "degenerate_count": 14,
+    "kl_divergence": null,
+    "spectral_certification": null
+  },
+  "kl_contributions": {},
+  "cot_preserved_layers": [],
+  "float_layer_weights": {},
+  "lora_adapters_saved": false
+}

config.json

@@ -0,0 +1,32 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dtype": "float16",
+  "eos_token_id": 2,
+  "head_dim": 64,
+  "hidden_act": "silu",
+  "hidden_size": 512,
+  "initializer_range": 0.02,
+  "intermediate_size": 1408,
+  "max_position_embeddings": 1024,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "num_attention_heads": 8,
+  "num_hidden_layers": 12,
+  "num_key_value_heads": 4,
+  "pad_token_id": 1,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 10000,
+    "rope_type": "default"
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.8.1",
+  "use_cache": false,
+  "vocab_size": 32000
+}

generation_config.json

@@ -0,0 +1,11 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": 2,
+  "max_length": 20,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "pad_token_id": 1,
+  "transformers_version": "5.8.1",
+  "use_cache": true
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c758e5d5ac86aac093a6df0c836a244e9ca800f1006fa4f33993891d5e3bc828
+size 103584504

obliteratus_session.json

@@ -0,0 +1,9 @@
+{
+  "label": "rdo on Supra-50M-Reasoning (14:30)",
+  "model_id": "SupraLabs/Supra-50M-Reasoning",
+  "model_choice": "SupraLabs/Supra-50M-Reasoning",
+  "method": "rdo",
+  "dataset_key": "anthropic_redteam",
+  "prompt_volume": -1,
+  "source": "obliterate"
+}

tokenizer_config.json

@@ -0,0 +1,18 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<s>",
+  "eos_token": "</s>",
+  "is_local": true,
+  "local_files_only": false,
+  "max_length": 256,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_to_multiple_of": null,
+  "pad_token": "<pad>",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "stride": 0,
+  "tokenizer_class": "TokenizersBackend",
+  "truncation_side": "right",
+  "truncation_strategy": "longest_first",
+  "unk_token": "<unk>"
+}