mirror of
https://github.com/elder-plinius/OBLITERATUS.git
synced 2026-07-12 23:26:32 +02:00
faber
04b8ec60cb
New core modules: - auto_obliterate.py: Automated multi-iteration obliteration pipeline - watchtower.py: HF Hub model discovery and tracking - ui_watchtower.py: Gradio tabs for Watchtower (ready for app.py wiring) - hard_negative.py: Residue mining from refusal audits - model_profile.py: Parameter profiling from safetensors/config - bestiary_sync.py: Sync models from PlinyOS BESTIARY registry - models_client.py: Lightweight HF model list client Framework enhancements: - abliterate.py: ASPA source-tethering, step gradient blending, hard-negative residue support - cli.py: self-improve command, model profiling, hard-negative flags - prompts.py: Expanded 842-prompt refusal eval corpus across 10 categories - __init__.py: New exports (Watchtower, AutoObliterator) Reference implementations (14 scripts): - ASPA sweep, gradient search, coherence eval, MMLU benchmarks - Pareto controller, refusal sniper, stock comparisons Documentation: - README: Research framing, responsible use section, comprehensive disclaimer - docs/beyond_sota_roadmap.md, docs/recursive_self_improvement.md Tests: 4 new test files (354 lines) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
79 lines
2.3 KiB
Python
79 lines
2.3 KiB
Python
from __future__ import annotations
|
|
|
|
from types import SimpleNamespace
|
|
|
|
import pytest
|
|
import torch.nn as nn
|
|
|
|
from obliteratus.models import loader
|
|
from obliteratus.models.loader import ModelHandle
|
|
from obliteratus.strategies.utils import (
|
|
get_attention_module,
|
|
get_embedding_module,
|
|
get_ffn_module,
|
|
get_layer_modules,
|
|
)
|
|
|
|
|
|
class _Gemma4LikeLayer(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.self_attn = nn.Module()
|
|
self.mlp = nn.Module()
|
|
|
|
|
|
class _Gemma4LikeModel(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.model = nn.Module()
|
|
self.model.language_model = nn.Module()
|
|
self.model.language_model.layers = nn.ModuleList(
|
|
[_Gemma4LikeLayer(), _Gemma4LikeLayer()]
|
|
)
|
|
self.model.language_model.embed_tokens = nn.Embedding(16, 8)
|
|
self.lm_head = nn.Linear(8, 16, bias=False)
|
|
|
|
|
|
def _gemma4_config(model_type: str = "gemma4_unified"):
|
|
return SimpleNamespace(
|
|
model_type=model_type,
|
|
architectures=["Gemma4UnifiedForConditionalGeneration"],
|
|
text_config=SimpleNamespace(
|
|
num_hidden_layers=2,
|
|
num_attention_heads=4,
|
|
hidden_size=8,
|
|
intermediate_size=32,
|
|
),
|
|
)
|
|
|
|
|
|
def test_gemma4_unified_uses_image_text_auto_model_when_available():
|
|
if loader.AutoModelForImageTextToText is None:
|
|
with pytest.raises(RuntimeError, match="AutoModelForImageTextToText"):
|
|
loader._select_model_class("causal_lm", _gemma4_config())
|
|
return
|
|
|
|
assert (
|
|
loader._select_model_class("causal_lm", _gemma4_config())
|
|
is loader.AutoModelForImageTextToText
|
|
)
|
|
|
|
|
|
def test_gemma4_nested_language_stack_paths():
|
|
handle = ModelHandle(
|
|
model=_Gemma4LikeModel(),
|
|
tokenizer=SimpleNamespace(pad_token="<pad>", eos_token="<eos>"),
|
|
config=_gemma4_config(),
|
|
model_name="google/gemma-4-12B-it",
|
|
task="causal_lm",
|
|
)
|
|
|
|
layers = get_layer_modules(handle)
|
|
|
|
assert len(layers) == 2
|
|
assert get_attention_module(layers[0], handle.architecture) is layers[0].self_attn
|
|
assert get_ffn_module(layers[0], handle.architecture) is layers[0].mlp
|
|
assert get_embedding_module(handle).num_embeddings == 16
|
|
assert handle.num_layers == 2
|
|
assert handle.hidden_size == 8
|