vh/heretic
1
0
Fork 0
mirror of https://github.com/p-e-w/heretic.git synced 2026-10-01 16:31:28 -07:00
Code Issues Packages Projects Releases Wiki Activity

feat: modifier plugins + ARA (#446)

Browse Source 
* 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>
This commit is contained in:
Philipp Emanuel Weidmann
2026-10-01 14:30:07 +05:30
committed by GitHub
co-authored by kabachuha joninco Ashar
parent 71e6d5eb38
commit 662e4ba27e
33 changed files with 1944 additions and 1738 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/gemma-4e/SHA256SUMS.ci3
+7
View File
@@ -0,0 +1,7 @@
2f1b4d75d067bae3fe44e676721c7f077d243bc007156cb9c2f8b5836613d082 *chat_template.jinja
ca80080dfa4ec6ba87152fa2b9afe70b90c400e5c4b1d6bdc3aa3114467ca68f *config.json
70070bac883cf9c39b5992450d6b23cd160eaf33099e24c654e0359d2f87c760 *generation_config.json
9ff0593e3fbd0ba463bbc980ebb3ed34798e562b606f90cd93c2df5403732c7b *model.safetensors
32bdf45d2ad4cc29a0822ddd157a182de76644f0419a6228d151495256e9813c *processor_config.json
cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f *tokenizer.json
a1bab8c81ed15fa6ce912ec993c66cb49392e0487fb1ea5f5f11ea3618683627 *tokenizer_config.json
tests/gemma-4e/config.toml
+20 -3
View File
@@ -1,4 +1,4 @@
# This test case is for Hybrid-Edge models.
# This test case is for hybrid models.
# After any change related to it, this test should PASS.
model = "tiny-random/gemma-4e"
@@ -9,6 +9,11 @@ print_debug_information = true
batch_size = 2
max_response_length = 10
modifiers = [
{ plugin = "heretic.modifiers.abliteration.Abliteration" },
]
n_trials = 2
n_startup_trials = 1
@@ -18,13 +23,13 @@ trial_index = 0
model_action = "save"
save_directory = "model"
[good_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[bad_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
@@ -41,3 +46,15 @@ dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "test[:5]"
column = "text"
[modifier.Abliteration.good_prompts]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[modifier.Abliteration.bad_prompts]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
column = "text"
tests/gpt-oss/SHA256SUMS.linux
+6
View File
@@ -0,0 +1,6 @@
f8d9255777615591a7cc1a7c932f5a69e181128902295e1b81221d20d983cac7 *chat_template.jinja
91d2a5190c7ea0f74ed499428d4ad62b5208d63b36f7bbb562d15f4be25bd5c2 *config.json
dd6034a30113decdfaf8886622e20eb9e2e02d3f774918d474a4e26cfd7fbba8 *generation_config.json
aefe8b9c4b4969f6d13c5d778760f3dce4e25134324b33677934550d9df02a7c *model.safetensors
fce342a4642cb8afc42d8d89cfa21198b64a43458ded7f6ff28d1151a08c9cda *tokenizer.json
9ba5fa877168e24823cb583c55b4c2e4df0331f30084953c7cf07de294640384 *tokenizer_config.json
tests/gpt-oss/config.toml
+56
View File
@@ -0,0 +1,56 @@
# This test case is for ARA.
# After any change related to it, this test should PASS.
model = "tiny-random/gpt-oss"
model_commit = "02ba5c61f879b5a38a8b1f7a8e0409b8e1bb8f38"
seed = 12345
print_debug_information = true
batch_size = 2
max_response_length = 10
n_trials = 2
n_startup_trials = 1
export_strategy = "merge"
checkpoint_action = "restart"
trial_index = 0
model_action = "save"
save_directory = "model"
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
column = "text"
[scorer.KLDivergence.prompts]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "test[:5]"
column = "text"
[scorer.KeywordRate.prompts]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "test[:5]"
column = "text"
[modifier.ARA.good_prompts]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[modifier.ARA.bad_prompts]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
column = "text"
tests/minicpm5/config.toml
+22 -4
View File
@@ -9,6 +9,11 @@ print_debug_information = true
batch_size = 2
max_response_length = 10
modifiers = [
{ plugin = "heretic.modifiers.abliteration.Abliteration" },
]
n_trials = 2
n_startup_trials = 1
@@ -18,15 +23,13 @@ trial_index = 0
model_action = "save"
save_directory = "model"
row_normalization = "none"
[good_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[bad_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
@@ -43,3 +46,18 @@ dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "test[:5]"
column = "text"
[modifier.Abliteration]
row_normalization = "none"
[modifier.Abliteration.good_prompts]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[modifier.Abliteration.bad_prompts]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
column = "text"
tests/mistral-3/config.toml
+20 -3
View File
@@ -1,4 +1,4 @@
# This test case is for Dense models.
# This test case is for dense models.
# After any change related to it, this test should PASS.
model = "tiny-random/mistral-3"
@@ -9,6 +9,11 @@ print_debug_information = true
batch_size = 2
max_response_length = 10
modifiers = [
{ plugin = "heretic.modifiers.abliteration.Abliteration" },
]
n_trials = 2
n_startup_trials = 1
@@ -18,13 +23,13 @@ trial_index = 0
model_action = "save"
save_directory = "model"
[good_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[bad_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
@@ -41,3 +46,15 @@ dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "test[:5]"
column = "text"
[modifier.Abliteration.good_prompts]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[modifier.Abliteration.bad_prompts]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
column = "text"
tests/qwen2.5/config.toml
+22 -4
View File
@@ -9,6 +9,11 @@ print_debug_information = true
batch_size = 2
max_response_length = 10
modifiers = [
{ plugin = "heretic.modifiers.abliteration.Abliteration" },
]
n_trials = 2
n_startup_trials = 1
@@ -18,15 +23,13 @@ trial_index = 0
model_action = "save"
save_directory = "model"
row_normalization = "pre"
[good_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[bad_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
@@ -43,3 +46,18 @@ dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "test[:5]"
column = "text"
[modifier.Abliteration]
row_normalization = "pre"
[modifier.Abliteration.good_prompts]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[modifier.Abliteration.bad_prompts]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
column = "text"
tests/qwen3.5-moe/config.toml
+19 -2
View File
@@ -9,6 +9,11 @@ print_debug_information = true
batch_size = 2
max_response_length = 10
modifiers = [
{ plugin = "heretic.modifiers.abliteration.Abliteration" },
]
n_trials = 2
n_startup_trials = 1
@@ -18,13 +23,13 @@ trial_index = 0
model_action = "save"
save_directory = "model"
[good_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[bad_prompts]
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
@@ -41,3 +46,15 @@ dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "test[:5]"
column = "text"
[modifier.Abliteration.good_prompts]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[modifier.Abliteration.bad_prompts]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
column = "text"
tests/seed-oss/SHA256SUMS.linux
+6
View File
@@ -0,0 +1,6 @@
39a23a1a28f68acd747373fa688bea630021ec4db2bc8dec2fd3962240fb0418 *chat_template.jinja
35e3ad5aeaa78984629ceba6c22a4236b4c21d6959485900da2a5a98ea7e4062 *config.json
2ec7a17f55287d482ca956f5acff44d9e88f6c7fd5ec06c7c33d641afa9af330 *generation_config.json
0cdea9064dcbe6db666f9d42a283d664c133d4c67a2f6ea0fd863a54f522160c *model.safetensors
3cf3a6d9520f195638a36f0194239d817de7288710bca55f1f5753de226748f7 *tokenizer.json
388b47e61cb40f2fd51a89999053686ab4c45b40b43c0329d15645f5910d069e *tokenizer_config.json
tests/seed-oss/config.toml
+65
View File
@@ -0,0 +1,65 @@
# This test case is for ARA (non-standard settings).
# After any change related to it, this test should PASS.
model = "tiny-random/seed-oss"
model_commit = "6860befd78b678885f7a52bbf41d7fd0671af2db"
seed = 12345
print_debug_information = true
batch_size = 2
max_response_length = 10
scorers = [
{ plugin = "heretic.scorers.keyword_rate.KeywordRate", optimization = "minimize" },
{ plugin = "heretic.scorers.kl_divergence.KLDivergence", optimization = "maximize" },
]
n_trials = 2
n_startup_trials = 1
export_strategy = "merge"
checkpoint_action = "restart"
trial_index = 0
model_action = "save"
save_directory = "model"
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[[response_prefix_test_prompts]]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
column = "text"
[scorer.KLDivergence.prompts]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "test[:5]"
column = "text"
[scorer.KeywordRate.prompts]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "test[:5]"
column = "text"
[modifier.ARA]
preserve_row_magnitudes = false
lora_rank = 20
[modifier.ARA.good_prompts]
dataset = "mlabonne/harmless_alpaca"
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
split = "train[:5]"
column = "text"
[modifier.ARA.bad_prompts]
dataset = "mlabonne/harmful_behaviors"
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
split = "train[:5]"
column = "text"