chore: add yet more CI test hashes

* feat: add modifier base class

* feat: re-implement abliteration as a modifier plugin

* fix: adjust good/bad prompts hack to match tests

* feat: support dataset specifications containing multiple individual datasets

* feat: re-implement ARA as a modifier plugin

Arbitrary-Rank Ablation (ARA) (Weidmann 2026) was originally introduced by @p-e-w in #211

Co-authored-by: kabachuha <artemkhrapov2001@yandex.ru>
Co-authored-by: joninco <joninco@bullpoint.org>
Co-authored-by: Ashar <coder3101@users.noreply.github.com>

* feat: add tests for ARA

* feat: reduce default number of trials

* ci: remove broken "semantic-pull-request" workflow

* ci: add yet another alternative model hash

---------

Co-authored-by: kabachuha <artemkhrapov2001@yandex.ru>
Co-authored-by: joninco <joninco@bullpoint.org>
Co-authored-by: Ashar <coder3101@users.noreply.github.com>

pyproject.toml

@@ -14,7 +14,6 @@ classifiers = [
     "Environment :: Console",
     "Environment :: GPU",
     "Intended Audience :: Science/Research",
-    "License :: OSI Approved :: GNU Affero General Public License v3 or later (AGPLv3+)",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
     "Programming Language :: Python :: 3",
     "Programming Language :: Python :: 3.10",
@@ -22,32 +21,32 @@ classifiers = [
     "Programming Language :: Python :: 3.12",
 ]
 dependencies = [
-    "accelerate~=1.13",
-    "bitsandbytes~=0.49",
-    "datasets~=4.7",
-    "huggingface-hub~=1.7",
+    "accelerate~=1.15",
+    "bitsandbytes~=0.50",
+    "datasets~=5.0",
+    "huggingface-hub~=1.33",
     "immutabledict~=4.3",
     "langdetect~=1.0",
     "lm-eval[hf]~=0.4",
     "numpy~=2.2",
-    "optuna~=4.7",
-    "peft~=0.19",
+    "optuna~=5.0",
+    "peft~=0.21",
     "psutil~=7.2",
     "py-cpuinfo~=9.0",
-    "pydantic-settings~=2.13",
+    "pydantic-settings~=2.15",
     "questionary~=2.1",
-    "rich~=14.3",
+    "rich~=15.0",
     "tomli-w~=1.2",
     "torch",  # version deliberately unspecified
     "torchvision",  # version deliberately unspecified
-    "tqdm~=4.67",
-    "transformers[kernels]~=5.6",
+    "tqdm~=4.70",
+    "transformers[kernels]~=5.18",
 ]
 
 [dependency-groups]
 dev = [
-    "ruff>=0.14.5",
-    "ty>=0.0.5",
+    "ruff>=0.16.9",
+    "ty>=0.0.84",
 ]
 
 [project.urls]
@@ -61,11 +60,31 @@ Changelog = "https://github.com/p-e-w/heretic/releases"
 heretic = "heretic.main:main"
 
 [build-system]
-requires = ["uv_build>=0.8.11,<0.9.0"]
+requires = ["uv_build>=0.12.21,<0.13.0"]
 build-backend = "uv_build"
 
+[tool.ruff.lint]
+ignore = [
+    "B023",  # "Function definition does not bind loop variable"
+    "BLE001",  # "Do not catch blind exception"
+    "TRY002",  # "Create your own exception"
+    "SIM117",  # "Use a single `with` statement with multiple contexts instead of nested `with` statements"
+]
+
 [tool.uv]
-exclude-newer = "7 days"
+# TODO: Re-enable before 2.0 release!
+#exclude-newer = "7 days"
+
+[tool.uv.sources]
+torch = { index = "pytorch" }
+torchvision = { index = "pytorch" }
+
+[[tool.uv.index]]
+# Default to CPU wheels for PyTorch to avoid issues in CI.
+# Can be overridden by passing `--index pytorch=<url>` to uv.
+name = "pytorch"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true
 
 [tool.uv.build-backend]
 module-name = "heretic"

src/heretic/config.py

@@ -2,7 +2,7 @@
 # Copyright (C) 2025-2026  Philipp Emanuel Weidmann <pew@worldwidemann.com> + contributors
 
 from enum import Enum
-from typing import Dict, Literal, TypeAlias
+from typing import Literal, TypeAlias
 
 from pydantic import (
     BaseModel,
@@ -254,12 +254,12 @@ class Settings(BaseSettings):
         ),
     )
 
-    device_map: str | Dict[str, int | str] = Field(
+    device_map: str | dict[str, int | str] = Field(
         default="auto",
         description="Device map to pass to Accelerate when loading the model.",
     )
 
-    max_memory: Dict[str, str] | None = Field(
+    max_memory: dict[str, str] | None = Field(
         default=None,
         description='Maximum memory to allocate per device (e.g., { "0" = "20GB", "cpu" = "64GB" }).',
     )

src/heretic/main.py

@@ -1,8 +1,6 @@
 # SPDX-License-Identifier: AGPL-3.0-or-later
 # Copyright (C) 2025-2026  Philipp Emanuel Weidmann <pew@worldwidemann.com> + contributors
 
-# ruff: noqa: E402
-
 import sys
 
 # Ensure standard output/error use UTF-8 instead of system default charmap (e.g. cp1252 on Windows).
@@ -23,7 +21,7 @@ def _is_help_invocation() -> bool:
 
 # Parse and handle CLI help before importing heavyweight ML/runtime dependencies.
 if _is_help_invocation():
-    Settings()  # ty:ignore[missing-argument]
+    Settings()
 
 # FIXME: Rich progress bars are currently disabled because of rendering issues
 #        when used from multiple threads in parallel (e.g. by huggingface_hub).
@@ -223,7 +221,7 @@ def run():
     try:
         # The required argument "model" must be provided by the user,
         # either on the command line or in the configuration file.
-        settings = Settings()  # ty:ignore[missing-argument]
+        settings = Settings()
     except ValidationError as error:
         print(f"[red]Configuration contains [bold]{error.error_count()}[/] errors:[/]")
 
@@ -251,13 +249,11 @@ def run():
 
         if reproduction_information["version"] != "4":
             print(
-                (
                 f"[red]Unsupported file format version: [bold]{reproduction_information['version']}[/].[/] "
                 "This version of Heretic reads version 4 (plugin-based) reproduce.json files. "
                 "Older files were produced before the introduction of the plugin system and are not supported. "
                 "Please install Heretic 1.4 to use these files."
             )
-            )
             return
 
         if not check_environment(settings, reproduction_information):
@@ -339,13 +335,11 @@ def run():
             if settings.checkpoint_action is None:
                 print()
                 print(
-                    (
                     "[green]You have already processed this model.[/] "
                     "You can show the results from the previous run, allowing you to export models or to run additional trials. "
                     "Alternatively, you can ignore the previous run and start from scratch. "
                     "This will delete the checkpoint file and all results from the previous run."
                 )
-                )
 
             choices.append(
                 Choice(
@@ -357,13 +351,11 @@ def run():
             if settings.checkpoint_action is None:
                 print()
                 print(
-                    (
                     "[yellow]You have already processed this model, but the run was interrupted.[/] "
                     "You can continue the previous run from where it stopped. This will override any specified settings. "
                     "Alternatively, you can ignore the previous run and start from scratch. "
                     "This will delete the checkpoint file and all results from the previous run."
                 )
-                )
 
             choices.append(
                 Choice(
@@ -771,13 +763,11 @@ def run():
             if settings.trial_index is None:
                 print()
                 print(
-                    (
                     "The following trials resulted in Pareto optimal combinations of the optimization objectives. "
                     "After selecting a trial, you will be able to save the model, upload it to Hugging Face, "
                     "chat with it to test how well it works, or run standard benchmarks on it. "
                     "You can return to this menu later to select a different trial. "
                 )
-                )
 
         while trial_loop_active:
             # Ensure a predefined trial is only processed once.
@@ -1077,14 +1067,12 @@ def run():
                             if is_reproducible:
                                 if settings.upload_reproducibility_information is None:
                                     print(
-                                        (
                                         "Heretic can add information to the repository that allows others to reproduce the model. "
                                         "This is optional, but valuable to the community as both a learning tool and to preserve computational work already done. "
                                         "Guaranteeing reproducibility requires basic system information (Python and OS version, CPU and GPU/accelerator info) "
                                         "as tensor operations can give different results in different system environments. "
                                         "[bold]The information does not include any file system paths or other private data.[/]"
                                     )
-                                    )
 
                                 reproducibility_information = ask_if_unset(
                                     settings.upload_reproducibility_information,
@@ -1115,7 +1103,7 @@ def run():
                             if strategy == ExportStrategy.ADAPTER:
                                 print("Uploading LoRA adapter...")
                                 model.model.push_to_hub(
-                                    repo_id,
+                                    repo_id,  # ty: ignore[invalid-argument-type]
                                     private=private,
                                     max_shard_size=settings.max_shard_size,
                                     token=token,
@@ -1124,7 +1112,7 @@ def run():
                                 print("Uploading merged model...")
                                 merged_model = model.get_merged_model()
                                 merged_model.push_to_hub(
-                                    repo_id,
+                                    repo_id,  # ty: ignore[invalid-argument-type]
                                     private=private,
                                     max_shard_size=settings.max_shard_size,
                                     token=token,
@@ -1319,7 +1307,7 @@ def run():
                             benchmark_original_model = scope == "Benchmark both models"
 
                             hflm = HFLM(
-                                pretrained=model.model,  # ty:ignore[invalid-argument-type]
+                                pretrained=model.model,
                                 tokenizer=model.tokenizer,  # ty:ignore[invalid-argument-type]
                                 batch_size="auto",
                             )

src/heretic/model.py

@@ -1,8 +1,9 @@
 # SPDX-License-Identifier: AGPL-3.0-or-later
 # Copyright (C) 2025-2026  Philipp Emanuel Weidmann <pew@worldwidemann.com> + contributors
 
+from collections.abc import Callable
 from contextlib import suppress
-from typing import Any, Callable, Type, TypeAlias, cast
+from typing import Any, TypeAlias, cast
 
 import torch
 from peft import LoraConfig, PeftModel, get_peft_model
@@ -23,7 +24,7 @@ from transformers import (
     TextStreamer,
 )
 from transformers.generation import (
-    GenerateDecoderOnlyOutput,  # ty:ignore[possibly-missing-import]
+    GenerateDecoderOnlyOutput,
 )
 
 from .config import QuantizationMethod, Settings
@@ -33,10 +34,10 @@ from .utils import Prompt, batchify, format_exception, print
 
 def get_model_class(
     model: str,
-) -> Type[AutoModelForImageTextToText] | Type[AutoModelForCausalLM]:
+) -> type[AutoModelForImageTextToText] | type[AutoModelForCausalLM]:
     configs = PretrainedConfig.get_config_dict(model)
 
-    if any([("vision_config" in config) for config in configs]):
+    if any(("vision_config" in config) for config in configs):
         return AutoModelForImageTextToText
     else:
         return AutoModelForCausalLM
@@ -68,9 +69,14 @@ class Model:
         print()
         print(f"Loading model [bold]{settings.model}[/]...")
 
-        self.tokenizer = AutoTokenizer.from_pretrained(
+        # PreTrainedTokenizerBase is the "base class for all tokenizer backends"
+        # according to the documentation.
+        self.tokenizer = cast(
+            PreTrainedTokenizerBase,
+            AutoTokenizer.from_pretrained(
                 settings.model,
                 **self.revision_kwargs,
+            ),
         )
 
         # Multimodal models have a processor we'll want to save.
@@ -90,7 +96,7 @@ class Model:
         #           after the prompt and thinks the sequence is complete.
         self.tokenizer.padding_side = "left"
 
-        self.model = None  # ty:ignore[invalid-assignment]
+        self.model = None
         self.max_memory = (
             {int(k) if k.isdigit() else k: v for k, v in settings.max_memory.items()}
             if settings.max_memory
@@ -143,7 +149,7 @@ class Model:
                     max_new_tokens=1,
                 )
             except Exception as error:
-                self.model = None  # ty:ignore[invalid-assignment]
+                self.model = None
                 empty_cache()
 
                 formatted = format_exception(error)
@@ -315,7 +321,7 @@ class Model:
             return True
 
         # Purge existing model object from memory to make space.
-        self.model = None  # ty:ignore[invalid-assignment]
+        self.model = None
         empty_cache()
 
         quantization_config = self._get_quantization_config(
@@ -352,10 +358,10 @@ class Model:
 
         # Most multimodal models.
         with suppress(Exception):
-            return model.model.language_model.layers
+            return model.model.language_model.layers  # ty: ignore[unresolved-attribute, invalid-return-type]
 
         # Text-only models.
-        return model.model.layers
+        return model.model.layers  # ty: ignore[unresolved-attribute, invalid-return-type]
 
     def get_layer_modules(self, layer_index: int) -> dict[str, list[Module]]:
         layer = self.get_layers()[layer_index]
@@ -376,50 +382,50 @@ class Model:
 
         # Standard self-attention out-projection (most models).
         with suppress(Exception):
-            try_add("attn.o_proj", layer.self_attn.o_proj)  # ty:ignore[possibly-missing-attribute]
+            try_add("attn.o_proj", layer.self_attn.o_proj)  # ty: ignore[unresolved-attribute]
 
         # Qwen3.5 MoE hybrid layers use GatedDeltaNet (linear attention) instead of
         # standard self-attention, so self_attn.o_proj doesn't exist on those layers.
         with suppress(Exception):
-            try_add("attn.o_proj", layer.linear_attn.out_proj)  # ty:ignore[possibly-missing-attribute]
+            try_add("attn.o_proj", layer.linear_attn.out_proj)  # ty: ignore[unresolved-attribute]
 
         # Most dense models.
         with suppress(Exception):
-            try_add("mlp.down_proj", layer.mlp.down_proj)  # ty:ignore[possibly-missing-attribute]
+            try_add("mlp.down_proj", layer.mlp.down_proj)  # ty: ignore[unresolved-attribute]
 
         # Some MoE models (e.g. Qwen3).
         with suppress(Exception):
-            for expert in layer.mlp.experts:  # ty:ignore[possibly-missing-attribute, not-iterable]
-                try_add("mlp.down_proj", expert.down_proj)  # ty:ignore[possibly-missing-attribute]
+            for expert in layer.mlp.experts:  # ty:ignore[not-iterable, unresolved-attribute]
+                try_add("mlp.down_proj", expert.down_proj)  # ty: ignore[unresolved-attribute]
 
         # Phi-3.5-MoE (and possibly others).
         with suppress(Exception):
-            for expert in layer.block_sparse_moe.experts:  # ty:ignore[possibly-missing-attribute, not-iterable]
-                try_add("mlp.down_proj", expert.w2)  # ty:ignore[possibly-missing-attribute]
+            for expert in layer.block_sparse_moe.experts:  # ty:ignore[not-iterable, unresolved-attribute]
+                try_add("mlp.down_proj", expert.w2)  # ty: ignore[unresolved-attribute]
 
         # LFM dense operator blocks.
         with suppress(Exception):
-            try_add("attn.o_proj", layer.conv.out_proj)  # ty:ignore[possibly-missing-attribute]
+            try_add("attn.o_proj", layer.conv.out_proj)  # ty: ignore[unresolved-attribute]
 
         with suppress(Exception):
-            try_add("mlp.down_proj", layer.feed_forward.w2)  # ty:ignore[possibly-missing-attribute]
+            try_add("mlp.down_proj", layer.feed_forward.w2)  # ty: ignore[unresolved-attribute]
 
         # LFM transformer blocks.
         with suppress(Exception):
-            try_add("attn.o_proj", layer.self_attn.out_proj)  # ty:ignore[possibly-missing-attribute]
+            try_add("attn.o_proj", layer.self_attn.out_proj)  # ty: ignore[unresolved-attribute]
 
         with suppress(Exception):
-            for expert in layer.feed_forward.experts:  # ty:ignore[possibly-missing-attribute, not-iterable]
-                try_add("mlp.down_proj", expert.w2)  # ty:ignore[possibly-missing-attribute]
+            for expert in layer.feed_forward.experts:  # ty:ignore[not-iterable, unresolved-attribute]
+                try_add("mlp.down_proj", expert.w2)  # ty: ignore[unresolved-attribute]
 
         # Granite MoE Hybrid - attention layers with shared_mlp.
         with suppress(Exception):
-            try_add("mlp.down_proj", layer.shared_mlp.output_linear)  # ty:ignore[possibly-missing-attribute]
+            try_add("mlp.down_proj", layer.shared_mlp.output_linear)  # ty: ignore[unresolved-attribute]
 
         # Granite MoE Hybrid - MoE layers with experts.
         with suppress(Exception):
-            for expert in layer.moe.experts:  # ty:ignore[possibly-missing-attribute, not-iterable]
-                try_add("mlp.down_proj", expert.output_linear)  # ty:ignore[possibly-missing-attribute]
+            for expert in layer.moe.experts:  # ty:ignore[not-iterable, unresolved-attribute]
+                try_add("mlp.down_proj", expert.output_linear)  # ty: ignore[unresolved-attribute]
 
         # We need at least one module across all components for abliteration to work.
         total_modules = sum(len(mods) for mods in modules.values())
@@ -512,11 +518,12 @@ class Model:
         responses = []
 
         for batch in batchify(prompts, self.settings.batch_size):
-            for response in self.get_responses(
+            responses.extend(
+                self.get_responses(
                     batch,
                     skip_special_tokens=skip_special_tokens,
-            ):
-                responses.append(response)
+                )
+            )
 
         return responses
 
@@ -809,7 +816,7 @@ class Model:
             # The TextStreamer constructor annotates this parameter with the AutoTokenizer
             # type, which makes no sense because AutoTokenizer is a factory class,
             # not a base class that tokenizers inherit from.
-            self.tokenizer,  # ty:ignore[invalid-argument-type]
+            self.tokenizer,
             skip_prompt=True,
             skip_special_tokens=True,
         )

src/heretic/modifier.py

@@ -32,7 +32,7 @@ class Serializable(Protocol):
 Parameters = TypeVar("Parameters", bound=Serializable)
 
 
-class Modifier(Generic[Parameters], Plugin, ABC):
+class Modifier(Plugin, ABC, Generic[Parameters]):
     """
     Abstract base class for modifier plugins.
 

src/heretic/modifiers/abliteration.py

@@ -6,7 +6,7 @@ from dataclasses import asdict, dataclass
 from enum import Enum
 from typing import Any, cast
 
-import bitsandbytes as bnb
+import bitsandbytes.functional as BNB_F
 import torch
 import torch.linalg as LA
 import torch.nn.functional as F
@@ -332,9 +332,7 @@ class Abliteration(Modifier[Parameters]):
                 weight_distribution = parameters.weight_distributions[component]
 
                 # Type inference fails here for some reason.
-                distance = cast(
-                    float, abs(layer_index - weight_distribution.max_weight_position)
-                )
+                distance = abs(layer_index - weight_distribution.max_weight_position)
 
                 # Don't orthogonalize layers that are more than
                 # min_weight_distance away from max_weight_position.
@@ -390,15 +388,10 @@ class Abliteration(Modifier[Parameters]):
                         W = base_weight.to(torch.float32)
                     else:
                         # 4-bit quantization.
-                        # This cast is always valid. Type inference fails here because the
-                        # bnb.functional module is not found by ty for some reason.
-                        W = cast(
-                            Tensor,
-                            bnb.functional.dequantize_4bit(  # ty:ignore[possibly-missing-attribute]
+                        W = BNB_F.dequantize_4bit(
                             base_weight.data,
                             quant_state,
-                            ).to(torch.float32),
-                        )
+                        ).to(torch.float32)
 
                     # Flatten weight matrix to (out_features, in_features).
                     W = W.view(W.shape[0], -1)

src/heretic/modifiers/ara.py

@@ -7,7 +7,7 @@
 from dataclasses import asdict, dataclass
 from typing import Any, cast
 
-import bitsandbytes as bnb
+import bitsandbytes.functional as BNB_F
 import torch
 import torch.linalg as LA
 import torch.nn.functional as F
@@ -264,13 +264,10 @@ class ARA(Modifier[Parameters]):
                         W_base = base_weight.to(torch.float32)
                     else:
                         # Use the original dequantization logic from bitsandbytes.
-                        W_base = cast(
-                            Tensor,
-                            bnb.functional.dequantize_4bit(  # ty:ignore[possibly-missing-attribute]
+                        W_base = BNB_F.dequantize_4bit(
                             base_weight.data,
                             quant_state,
-                            ).to(torch.float32),
-                        )
+                        ).to(torch.float32)
 
                     # Pre-calculate the original row norms to preserve them.
                     # See https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration

src/heretic/plugin.py

@@ -362,4 +362,4 @@ class Plugin:
         Override this in subclasses to do one-time setup (e.g. load prompts, compute
         baselines).
         """
-        return None
+        return

src/heretic/reproduce.py

@@ -81,7 +81,7 @@ def collect_reproducibles(path: str):
 
             found += 1
 
-            commit_hash = paths_info[0].last_commit.oid
+            commit_hash = paths_info[0].last_commit.oid  # ty: ignore[unresolved-attribute]
 
             file_path = (
                 Path(path)
@@ -285,13 +285,11 @@ def check_environment(
 
     else:
         print(
-            (
             "[yellow]The provided JSON file does not contain system information. "
             "Some system parameters can affect reproducibility, but due to the lack of system information, "
             "Heretic is unable to verify that those parameters match the original environment. "
             "Reproduction may or may not produce a byte-for-byte identical model.[/]"
         )
-        )
 
     requirements = get_requirements_dict()
     requirements["heretic-llm"] = format_version_information(
@@ -321,11 +319,9 @@ def check_environment(
     if system_mismatches or package_mismatches:
         print()
         print(
-            (
             "[yellow]Your local environment doesn't perfectly match the environment "
             "used to produce the original model. The following components differ:[/]"
         )
-        )
 
     if system_mismatches:
         table = Table()
@@ -358,13 +354,11 @@ def check_environment(
     if system_mismatches or package_mismatches:
         print()
         print(
-            (
             f"There is a {cast(MismatchSeverity, mismatch_severity).__rich__()} chance "
             "that reproduction won't produce a byte-for-byte identical model. "
             "However, the resulting model will very likely still behave similarly "
             "to the original model."
         )
-        )
 
         if settings.ignore_mismatches is None:
             print()

src/heretic/scorers/benchmark_score.py

@@ -44,7 +44,7 @@ class BenchmarkScore(Scorer):
         model = ctx.get_model()
 
         self.hflm = HFLM(
-            pretrained=model.model,  # ty:ignore[invalid-argument-type]
+            pretrained=model.model,
             tokenizer=model.tokenizer,  # ty:ignore[invalid-argument-type]
             batch_size="auto",
         )

src/heretic/system.py

@@ -96,7 +96,7 @@ def get_amdgpu_driver_version() -> str | None:
             if os.path.exists(version_path):
                 with open(version_path, "r", encoding="utf-8") as f:
                     return f.read().strip()
-    except Exception:
+    except Exception:  # noqa: S110
         pass
 
     return None
@@ -249,7 +249,7 @@ def get_accelerator_info_dict() -> dict[str, Any]:
         info: dict[str, Any] = {
             "type": "ROCm" if is_rocm else "CUDA",
             "api_name": "HIP Version" if is_rocm else "CUDA Version",
-            "api_version": torch.version.hip if is_rocm else torch.version.cuda,  # ty:ignore[unresolved-attribute]
+            "api_version": torch.version.hip if is_rocm else torch.version.cuda,
             "driver_version": get_amdgpu_driver_version()
             if is_rocm
             else get_nvidia_driver_version(),
@@ -264,13 +264,13 @@ def get_accelerator_info_dict() -> dict[str, Any]:
         return info
 
     if is_xpu_available():
-        count = torch.xpu.device_count()  # ty:ignore[unresolved-attribute]
+        count = torch.xpu.device_count()
         return {
             "type": "XPU",
             "api_name": None,
             "api_version": None,
             "driver_version": get_xpu_driver_version(),
-            "devices": [{"name": torch.xpu.get_device_name(i)} for i in range(count)],  # ty:ignore[unresolved-attribute]
+            "devices": [{"name": torch.xpu.get_device_name(i)} for i in range(count)],
         }
 
     if is_mlu_available():

tests/gemma-4e/SHA256SUMS.ci

@@ -1,6 +1,6 @@
 2f1b4d75d067bae3fe44e676721c7f077d243bc007156cb9c2f8b5836613d082 *chat_template.jinja
-ca80080dfa4ec6ba87152fa2b9afe70b90c400e5c4b1d6bdc3aa3114467ca68f *config.json
-70070bac883cf9c39b5992450d6b23cd160eaf33099e24c654e0359d2f87c760 *generation_config.json
+c128bc8647a505e343c8cdff9bdd188b1a4a3f81938148685cbf90dff9827268 *config.json
+58678fb2b8ae1b96652dffa43f864ad1c1c59e49c889bab4264ba0844c4a23b8 *generation_config.json
 f3f4ec19504f182486459cf4e255ece265c25f827840d63b6a9d4058b8e4877a *model.safetensors
 32bdf45d2ad4cc29a0822ddd157a182de76644f0419a6228d151495256e9813c *processor_config.json
 cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f *tokenizer.json

tests/gemma-4e/SHA256SUMS.ci2

@@ -1,6 +1,6 @@
 2f1b4d75d067bae3fe44e676721c7f077d243bc007156cb9c2f8b5836613d082 *chat_template.jinja
-ca80080dfa4ec6ba87152fa2b9afe70b90c400e5c4b1d6bdc3aa3114467ca68f *config.json
-70070bac883cf9c39b5992450d6b23cd160eaf33099e24c654e0359d2f87c760 *generation_config.json
+c128bc8647a505e343c8cdff9bdd188b1a4a3f81938148685cbf90dff9827268 *config.json
+58678fb2b8ae1b96652dffa43f864ad1c1c59e49c889bab4264ba0844c4a23b8 *generation_config.json
 53c4ee891dce23c0ac85bebc2c4d48301469750fafbb3e6e024c15786d94db8b *model.safetensors
 32bdf45d2ad4cc29a0822ddd157a182de76644f0419a6228d151495256e9813c *processor_config.json
 cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f *tokenizer.json

tests/gemma-4e/SHA256SUMS.ci3

@@ -1,6 +1,6 @@
 2f1b4d75d067bae3fe44e676721c7f077d243bc007156cb9c2f8b5836613d082 *chat_template.jinja
-ca80080dfa4ec6ba87152fa2b9afe70b90c400e5c4b1d6bdc3aa3114467ca68f *config.json
-70070bac883cf9c39b5992450d6b23cd160eaf33099e24c654e0359d2f87c760 *generation_config.json
+c128bc8647a505e343c8cdff9bdd188b1a4a3f81938148685cbf90dff9827268 *config.json
+58678fb2b8ae1b96652dffa43f864ad1c1c59e49c889bab4264ba0844c4a23b8 *generation_config.json
 9ff0593e3fbd0ba463bbc980ebb3ed34798e562b606f90cd93c2df5403732c7b *model.safetensors
 32bdf45d2ad4cc29a0822ddd157a182de76644f0419a6228d151495256e9813c *processor_config.json
 cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f *tokenizer.json

tests/gemma-4e/SHA256SUMS.linux

@@ -1,6 +1,6 @@
 2f1b4d75d067bae3fe44e676721c7f077d243bc007156cb9c2f8b5836613d082 *chat_template.jinja
-ca80080dfa4ec6ba87152fa2b9afe70b90c400e5c4b1d6bdc3aa3114467ca68f *config.json
-70070bac883cf9c39b5992450d6b23cd160eaf33099e24c654e0359d2f87c760 *generation_config.json
+c128bc8647a505e343c8cdff9bdd188b1a4a3f81938148685cbf90dff9827268 *config.json
+58678fb2b8ae1b96652dffa43f864ad1c1c59e49c889bab4264ba0844c4a23b8 *generation_config.json
 effe36925f85ecb1e29bba84501a456bb49df21e4047be8b7ea3f6f88181fb65 *model.safetensors
 32bdf45d2ad4cc29a0822ddd157a182de76644f0419a6228d151495256e9813c *processor_config.json
 cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f *tokenizer.json

tests/gemma-4e/SHA256SUMS.windows

@@ -1,7 +0,0 @@
-b16d3228a775c549ba97af41233a54e9de8dd2b65250f78346661d18b936a8b5 *chat_template.jinja
-0094ad598a8043f84d82ad5c886547bca1d1d7f302d82f1491f83d388e89acd4 *config.json
-1a019c5d688d54cf01318eab88cb4345dfa52135eb1d83c2f54125469eb88d5c *generation_config.json
-effe36925f85ecb1e29bba84501a456bb49df21e4047be8b7ea3f6f88181fb65 *model.safetensors
-24d00232e58cfa179fe8b3911c788d4aad9a6279d778ebe4c72e82623b6197f9 *processor_config.json
-cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f *tokenizer.json
-8044bbbddaee8dc47e6b5660e013ba92224d4a5392b2939c59699aa0105f5c8b *tokenizer_config.json

tests/gpt-oss/SHA256SUMS.linux

@@ -1,6 +1,6 @@
 f8d9255777615591a7cc1a7c932f5a69e181128902295e1b81221d20d983cac7 *chat_template.jinja
-91d2a5190c7ea0f74ed499428d4ad62b5208d63b36f7bbb562d15f4be25bd5c2 *config.json
-dd6034a30113decdfaf8886622e20eb9e2e02d3f774918d474a4e26cfd7fbba8 *generation_config.json
+22d559dde87f1a0faf02efed3062af2b1fc8a1721a55b952cd5f5d8c356b4feb *config.json
+523e422e425e7a7ba23d85cc0faeef143dc5da37b72bfc649df5a8d83ee95beb *generation_config.json
 aefe8b9c4b4969f6d13c5d778760f3dce4e25134324b33677934550d9df02a7c *model.safetensors
 fce342a4642cb8afc42d8d89cfa21198b64a43458ded7f6ff28d1151a08c9cda *tokenizer.json
 9ba5fa877168e24823cb583c55b4c2e4df0331f30084953c7cf07de294640384 *tokenizer_config.json

tests/minicpm5/SHA256SUMS.linux

@@ -1,6 +1,6 @@
 7451a05cf1e28a79d97d7c0bc951028c0b1915119bf9046acd06a0e3d931f47c *chat_template.jinja
-fe6fd41d9f2ce5d6486748cf0330b574f37bf7d4e915f7b39d1af1a185cac3c3 *config.json
-c4c2ef5ae4a4e2dd10655a3b99d801a8a50497286ddd042ba35bcfefc44ad349 *generation_config.json
+de44f54b200f63c8ccb7899965bc9aaf21b23d35164c7de9a3cc3394530e3821 *config.json
+12d96814a0cd1a72cae14392b04b9644353e8600614faa6eb7ee5edbf9452ec1 *generation_config.json
 1535a9b7a91b2cb39ad280dbd9a940e2609a0b423d5b924df4d664e579912802 *model.safetensors
 ad92aaa8d3032c98a9158b8c5e8682bed10027ed6463e4fb1320fe5384210873 *tokenizer.json
 3ad32522c384dbe35192bb69de9befbf3f523e99d4bb3f95da757671d4c28281 *tokenizer_config.json

tests/minicpm5/SHA256SUMS.windows

@@ -1,6 +0,0 @@
-d8db3ff45c4c68a0ba9dee962ff1a0adde9a2be55e0895306f6bd2b2756f5adb *chat_template.jinja
-a9d6f64bb9d0c02b553119e475615153af625b5c2a16ccb8fb8b3c2cc348f465 *config.json
-0e7611a1e8fd0a06a139b0572b2c55b885ba9fb7db2022873c3508aebfb488aa *generation_config.json
-411d95f42d3e31aef41c28314c8f0431c980687a97904d32b4ef57c42199720f *model.safetensors
-ad92aaa8d3032c98a9158b8c5e8682bed10027ed6463e4fb1320fe5384210873 *tokenizer.json
-aa083f3da10340925734e876e41e235c459329294ecd35d7511ec5868c1f14e3 *tokenizer_config.json

tests/mistral-3/SHA256SUMS.ci

@@ -1,7 +1,7 @@
 39f03c383413f531fd302c06c7e982ad98c83f0657a8339ae25478ccb81fdcda *chat_template.jinja
-f69f84977a47c8fea9ce9fc26b7de379216cb01146ea726a87996d3554cfcd19 *config.json
-34dfa6012ca9ac5f57e5521d8dbaecbc7ab7f7ab0fd96ec020b543aab5f265d9 *generation_config.json
-876c6691eb85e3e5e11771e589529830fb454ab26344e1271ae550661e312b50 *model.safetensors
-84be30b124b50749c56d25fdbec5ccedf564446f6b3b035e88e1e07b986d2491 *processor_config.json
+b2cfd8a9da09efc97040f3839fe629f3925db856a44b58314301dc5e11345ec8 *config.json
+338d303c3c3df6884030cf5be24560f6e973fc183220d14acfb9364cd0e070e4 *generation_config.json
+05497f43a38427c5813105fc007ce0dd41be65470e9cb8f343443f31bab7f0fe *model.safetensors
+a05f93a41b7e42ccc18461b540fb54490c3ffa13aef79c800bbe7942e02e360f *processor_config.json
 c3a8d92e371b92a2cd6e678e31ebc27d0235e929a51fbf290f74742b341fa96f *tokenizer.json
 7b29c843c0043622d28fd4638451cbb0a609d99a0762ffbff3b92b4b2fee4d94 *tokenizer_config.json

tests/mistral-3/SHA256SUMS.ci2

@@ -1,7 +1,7 @@
 39f03c383413f531fd302c06c7e982ad98c83f0657a8339ae25478ccb81fdcda *chat_template.jinja
-f69f84977a47c8fea9ce9fc26b7de379216cb01146ea726a87996d3554cfcd19 *config.json
-34dfa6012ca9ac5f57e5521d8dbaecbc7ab7f7ab0fd96ec020b543aab5f265d9 *generation_config.json
-6febb813086f253e5ec0fcda02fdfc849c551a7dba54681b37ac5bc402e4eed6 *model.safetensors
-84be30b124b50749c56d25fdbec5ccedf564446f6b3b035e88e1e07b986d2491 *processor_config.json
+b2cfd8a9da09efc97040f3839fe629f3925db856a44b58314301dc5e11345ec8 *config.json
+338d303c3c3df6884030cf5be24560f6e973fc183220d14acfb9364cd0e070e4 *generation_config.json
+ad6675b44b476a899914761257bbeb7f320cb9c0dc39a7467daf4dad0216dd1f *model.safetensors
+a05f93a41b7e42ccc18461b540fb54490c3ffa13aef79c800bbe7942e02e360f *processor_config.json
 c3a8d92e371b92a2cd6e678e31ebc27d0235e929a51fbf290f74742b341fa96f *tokenizer.json
 7b29c843c0043622d28fd4638451cbb0a609d99a0762ffbff3b92b4b2fee4d94 *tokenizer_config.json

tests/mistral-3/SHA256SUMS.ci3

@@ -0,0 +1,7 @@
+39f03c383413f531fd302c06c7e982ad98c83f0657a8339ae25478ccb81fdcda *chat_template.jinja
+b2cfd8a9da09efc97040f3839fe629f3925db856a44b58314301dc5e11345ec8 *config.json
+338d303c3c3df6884030cf5be24560f6e973fc183220d14acfb9364cd0e070e4 *generation_config.json
+4cf4d463a94f477ed446e47130cb7da6b5994f0b788e75b82b3c65a531585621 *model.safetensors
+a05f93a41b7e42ccc18461b540fb54490c3ffa13aef79c800bbe7942e02e360f *processor_config.json
+c3a8d92e371b92a2cd6e678e31ebc27d0235e929a51fbf290f74742b341fa96f *tokenizer.json
+7b29c843c0043622d28fd4638451cbb0a609d99a0762ffbff3b92b4b2fee4d94 *tokenizer_config.json

tests/mistral-3/SHA256SUMS.linux

@@ -1,7 +1,7 @@
 39f03c383413f531fd302c06c7e982ad98c83f0657a8339ae25478ccb81fdcda *chat_template.jinja
-f69f84977a47c8fea9ce9fc26b7de379216cb01146ea726a87996d3554cfcd19 *config.json
-34dfa6012ca9ac5f57e5521d8dbaecbc7ab7f7ab0fd96ec020b543aab5f265d9 *generation_config.json
-29aff97d5633dead9e1ccd29a2cc153b4b7431d22f63c8d6cf60bc6547681cc9 *model.safetensors
-84be30b124b50749c56d25fdbec5ccedf564446f6b3b035e88e1e07b986d2491 *processor_config.json
+b2cfd8a9da09efc97040f3839fe629f3925db856a44b58314301dc5e11345ec8 *config.json
+338d303c3c3df6884030cf5be24560f6e973fc183220d14acfb9364cd0e070e4 *generation_config.json
+8244162ecb6ce7bad7b10fe4f015889d76a09c7749dc7ac8c4d957f0864b2e82 *model.safetensors
+a05f93a41b7e42ccc18461b540fb54490c3ffa13aef79c800bbe7942e02e360f *processor_config.json
 c3a8d92e371b92a2cd6e678e31ebc27d0235e929a51fbf290f74742b341fa96f *tokenizer.json
 7b29c843c0043622d28fd4638451cbb0a609d99a0762ffbff3b92b4b2fee4d94 *tokenizer_config.json

tests/mistral-3/SHA256SUMS.windows

@@ -1,7 +0,0 @@
-72f84af4ea36b82409c35e31b584361534305ef7c0d90fce20d0dc38a7efead8 *chat_template.jinja
-e4c5278b361c57621253c27a2c3db358e1580aec8a14be8e19d4420a224137cf *config.json
-8dde85c000ae807be907421465826c7c63a39f6acf6d04a5a84efaf116ed4ef7 *generation_config.json
-20b5a820b38438202c64e4fc9807bd19e29678bebd678d29b2ee2d2f5bf71587 *model.safetensors
-20e7a6dcde0a6f60ea3b4fb08f6f7afa62532dda93a3111e28384ba5150575f9 *processor_config.json
-c3a8d92e371b92a2cd6e678e31ebc27d0235e929a51fbf290f74742b341fa96f *tokenizer.json
-60a8042e29b4b20e884e48375aa1b9ac0025547371d50e60f6d55e6a9675e868 *tokenizer_config.json

tests/qwen2.5/SHA256SUMS.ci

@@ -1,6 +1,6 @@
 cd8e9439f0570856fd70470bf8889ebd8b5d1107207f67a5efb46e342330527f *chat_template.jinja
-45134b857367fdcb97c0179199848c353fc28f8b95ac2244ac8f45cca448d864 *config.json
-e81e23e025c38e825dcf8375861e26a90e804276e4db9ee390122a4fdc95dae7 *generation_config.json
-bd86541d817978c896bd3579e69ae6d41b6382eaf1646accf83d6feb16acb703 *model.safetensors
+6b517667960a5e39a692eb8277be7260ac02c22d9c7eda9c7aa1bbab21517fc1 *config.json
+64cf6f4e0016154cae45628221d77e3925961876125fe68bd942303ec7cca40f *generation_config.json
+e616cbeb5a913015eb3db96e001030048df2db560df363d4cf688f0c1b2c96de *model.safetensors
 f7f96da3a872b5e901575b2067c744ad336c3a3d77a21584d20024557b1bd7f0 *tokenizer.json
 04b1682c59acbd057f4c9072297faa73d56fc9de053094c659cdb4c464f58f86 *tokenizer_config.json

tests/qwen2.5/SHA256SUMS.linux

@@ -1,6 +1,6 @@
 cd8e9439f0570856fd70470bf8889ebd8b5d1107207f67a5efb46e342330527f *chat_template.jinja
-45134b857367fdcb97c0179199848c353fc28f8b95ac2244ac8f45cca448d864 *config.json
-e81e23e025c38e825dcf8375861e26a90e804276e4db9ee390122a4fdc95dae7 *generation_config.json
-e616cbeb5a913015eb3db96e001030048df2db560df363d4cf688f0c1b2c96de *model.safetensors
+6b517667960a5e39a692eb8277be7260ac02c22d9c7eda9c7aa1bbab21517fc1 *config.json
+64cf6f4e0016154cae45628221d77e3925961876125fe68bd942303ec7cca40f *generation_config.json
+bd86541d817978c896bd3579e69ae6d41b6382eaf1646accf83d6feb16acb703 *model.safetensors
 f7f96da3a872b5e901575b2067c744ad336c3a3d77a21584d20024557b1bd7f0 *tokenizer.json
 04b1682c59acbd057f4c9072297faa73d56fc9de053094c659cdb4c464f58f86 *tokenizer_config.json

tests/qwen2.5/SHA256SUMS.windows

@@ -1,6 +0,0 @@
-8aa40ce145adb73cb3a75194dc0224702a95850ec5275cabb728496bbd749fc6 *chat_template.jinja
-e8f2fcd2681eb92233c0902866441f79a207b235f0b03364d41ebf8c53df62a0 *config.json
-3fec6d7004e5ae311864de130b62e32dac87569874c91b3fe9c46e9309345c1c *generation_config.json
-bd86541d817978c896bd3579e69ae6d41b6382eaf1646accf83d6feb16acb703 *model.safetensors
-f7f96da3a872b5e901575b2067c744ad336c3a3d77a21584d20024557b1bd7f0 *tokenizer.json
-154e5ff1e7c152d964edf30da854ea62465c767719ac8e97e58babf2d4fa9079 *tokenizer_config.json

tests/qwen3.5-moe/SHA256SUMS.ci

@@ -1,7 +1,7 @@
 a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715 *chat_template.jinja
-749b56d1b1e08081981169db6f2c44ab0be4fd6ebb452d15baafa5e09c21586a *config.json
-4625d1d64d41d1fa9dae7af4ba1e1d7e65a194073d4efa58acb266a916eaaa74 *generation_config.json
-2b3e575ac065f11ae5d4a7c3740efccbed294b646f1645239191ee8393354e03 *model.safetensors
-01562eddd6f9e9ec4bc31656a3b7055284cafbf889acc6c4348dca431ae31f68 *processor_config.json
-87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4 *tokenizer.json
-2e31d1126e81bddf8d15c3f95260fb487b48c5131b24fcbb5bb9d2537e7afac0 *tokenizer_config.json
+4b501bef0727793de90d446fcfdc7757a9f8d19efc42ceaa3d26e90db4f876d5 *config.json
+532190619b0b243e0558a846257cc62883c26df812d456f8133e7b1043c69ad8 *generation_config.json
+cfa4d5428e9b23c245fd2413ad702688cbcacd25808ecb17889e97c0976532ac *model.safetensors
+3814737f2d4bfdd153a82364b3850f1be5dbc6f074a55c8d5fbf4d40127c3e87 *processor_config.json
+a5cd9732badce41de57e6efce8302930ded1c1188c5f81feb2bd6c24c4a1941f *tokenizer.json
+5e55dc6b9d9d28d49b6b2b80d2a2f2558a0b6e0e7b194bbca8664fc269ab3ebc *tokenizer_config.json

tests/qwen3.5-moe/SHA256SUMS.linux

@@ -1,7 +1,7 @@
 a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715 *chat_template.jinja
-749b56d1b1e08081981169db6f2c44ab0be4fd6ebb452d15baafa5e09c21586a *config.json
-4625d1d64d41d1fa9dae7af4ba1e1d7e65a194073d4efa58acb266a916eaaa74 *generation_config.json
-5e0fb0ac724cf079b693fc76a515e60bc16de72c32b36c107b9f078061c4f2ef *model.safetensors
-01562eddd6f9e9ec4bc31656a3b7055284cafbf889acc6c4348dca431ae31f68 *processor_config.json
-87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4 *tokenizer.json
-2e31d1126e81bddf8d15c3f95260fb487b48c5131b24fcbb5bb9d2537e7afac0 *tokenizer_config.json
+4b501bef0727793de90d446fcfdc7757a9f8d19efc42ceaa3d26e90db4f876d5 *config.json
+532190619b0b243e0558a846257cc62883c26df812d456f8133e7b1043c69ad8 *generation_config.json
+4c29b8ce99c59c90eca59aa6453029bda87165efa57cabc2c668599c02afc31a *model.safetensors
+3814737f2d4bfdd153a82364b3850f1be5dbc6f074a55c8d5fbf4d40127c3e87 *processor_config.json
+a5cd9732badce41de57e6efce8302930ded1c1188c5f81feb2bd6c24c4a1941f *tokenizer.json
+5e55dc6b9d9d28d49b6b2b80d2a2f2558a0b6e0e7b194bbca8664fc269ab3ebc *tokenizer_config.json

tests/qwen3.5-moe/SHA256SUMS.windows

@@ -1,7 +0,0 @@
-a92e1dd97cb1cb175c9b70c0828e146bea4371c2643319b661b777e89811972e *chat_template.jinja
-b75e911805663da79fb9fbbbcc917b8f1a285d2da54d95c2c63ea7c1ffe9a05a *config.json
-2cbd9df0e99570efcced23b8d777bdf1fc692efda54b21eb59ad56ade76c9db6 *generation_config.json
-6061519a9595326df41abcdd093892463793d4d026d6fd23548f1792f622a252 *model.safetensors
-0c29f9491e769aabbc389ad5912127cf6d9d5fceda2db8767f73d48131348c81 *processor_config.json
-87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4 *tokenizer.json
-4796e48d790a26d65f167bec8fc742beaa71f79f9468a6cd8b3ffa97f6e2a198 *tokenizer_config.json

tests/run_tests.py

@@ -53,7 +53,7 @@ for test_directory in script_directory.iterdir():
 
             print()
 
-            valid_hashes: dict[str, list[str]] = {}
+            valid_hashes: dict[str, set[str]] = {}
 
             for hash_file in hash_files:
                 with open(hash_file, "r", encoding="utf-8") as file:
@@ -63,26 +63,24 @@ for test_directory in script_directory.iterdir():
                             filename = filename.removeprefix("*")
 
                             if filename not in valid_hashes:
-                                valid_hashes[filename] = []
+                                valid_hashes[filename] = set()
 
-                            valid_hashes[filename].append(sha256.lower())
+                            valid_hashes[filename].add(sha256.lower())
 
             # Track which specific files failed within this test directory.
             failed_files: list[str] = []
-            for filename in valid_hashes:
+            for filename, hashes in valid_hashes.items():
                 sha256 = get_file_sha256(test_directory / "model" / filename)
 
-                if sha256.lower() not in valid_hashes[filename]:
+                if sha256.lower() not in hashes:
                     print(
-                        (
                         f"Test {test_directory.name} has FAILED!\n"
                         f"Output file {filename} doesn't match any valid hash.\n\n"
                         f"Valid hashes:\n"
-                            f"{chr(10).join(valid_hashes[filename])}\n\n"
+                        f"{chr(10).join(hashes)}\n\n"
                         f"Actual hash:\n"
                         f"{sha256}\n"
                     )
-                    )
                     failed_files.append(filename)
 
             if failed_files:

tests/seed-oss/SHA256SUMS.linux

@@ -1,6 +1,6 @@
 39a23a1a28f68acd747373fa688bea630021ec4db2bc8dec2fd3962240fb0418 *chat_template.jinja
-35e3ad5aeaa78984629ceba6c22a4236b4c21d6959485900da2a5a98ea7e4062 *config.json
-2ec7a17f55287d482ca956f5acff44d9e88f6c7fd5ec06c7c33d641afa9af330 *generation_config.json
+45d58c7316e9dab89e918b9ca1c141b8c7f5250d55e0f9be76f2f452903aeef8 *config.json
+02b35dff3057b998f224a3a5599dc7737ba028e3b4181ba89ba74ba1e707a5dc *generation_config.json
 0cdea9064dcbe6db666f9d42a283d664c133d4c67a2f6ea0fd863a54f522160c *model.safetensors
 3cf3a6d9520f195638a36f0194239d817de7288710bca55f1f5753de226748f7 *tokenizer.json
 388b47e61cb40f2fd51a89999053686ab4c45b40b43c0329d15645f5910d069e *tokenizer_config.json