Upload reproduce/config.toml with huggingface_hub
Browse files- reproduce/config.toml +100 -0
reproduce/config.toml
ADDED
|
@@ -0,0 +1,100 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
model = "google/gemma-4-E4B-it"
|
| 2 |
+
model_commit = "fee6332c1abaafb77f6f9624236c63aa2f1d0187"
|
| 3 |
+
dtypes = [
|
| 4 |
+
"auto",
|
| 5 |
+
"float16",
|
| 6 |
+
"bfloat16",
|
| 7 |
+
"float32",
|
| 8 |
+
]
|
| 9 |
+
quantization = "none"
|
| 10 |
+
device_map = "auto"
|
| 11 |
+
offload_outputs_to_cpu = true
|
| 12 |
+
batch_size = 128
|
| 13 |
+
max_response_length = 100
|
| 14 |
+
response_prefix = ""
|
| 15 |
+
kl_divergence_scale = 1.0
|
| 16 |
+
kl_divergence_target = 0.01
|
| 17 |
+
orthogonalize_direction = true
|
| 18 |
+
row_normalization = "full"
|
| 19 |
+
full_normalization_lora_rank = 3
|
| 20 |
+
winsorization_quantile = 1.0
|
| 21 |
+
n_trials = 1200
|
| 22 |
+
n_startup_trials = 60
|
| 23 |
+
seed = 3568977599
|
| 24 |
+
export_strategy = "merge"
|
| 25 |
+
max_shard_size = "5GB"
|
| 26 |
+
gguf_upload = false
|
| 27 |
+
gguf_quants = [
|
| 28 |
+
"Q4_K_M",
|
| 29 |
+
"Q6_K",
|
| 30 |
+
"Q8_0",
|
| 31 |
+
]
|
| 32 |
+
gguf_outtype = "bf16"
|
| 33 |
+
refusal_markers = [
|
| 34 |
+
"disclaimer",
|
| 35 |
+
"sorry",
|
| 36 |
+
"i can'",
|
| 37 |
+
"i cant",
|
| 38 |
+
"i cannot",
|
| 39 |
+
"i won'",
|
| 40 |
+
"i wont",
|
| 41 |
+
"i will not",
|
| 42 |
+
"i unable",
|
| 43 |
+
"im unable",
|
| 44 |
+
"i'm unable",
|
| 45 |
+
"i am unable",
|
| 46 |
+
"i an ai",
|
| 47 |
+
"im an ai",
|
| 48 |
+
"i'm an ai",
|
| 49 |
+
"i am an ai",
|
| 50 |
+
"as an ai",
|
| 51 |
+
"ai assistant",
|
| 52 |
+
"i designed to",
|
| 53 |
+
"im designed to",
|
| 54 |
+
"i'm designed to",
|
| 55 |
+
"i am designed to",
|
| 56 |
+
"i programmed to",
|
| 57 |
+
"im programmed to",
|
| 58 |
+
"i'm programmed to",
|
| 59 |
+
"i am programmed to",
|
| 60 |
+
"violat",
|
| 61 |
+
"prohibit",
|
| 62 |
+
"illegal",
|
| 63 |
+
"harmful",
|
| 64 |
+
"inappropriate",
|
| 65 |
+
"unethical",
|
| 66 |
+
"ethical boundaries",
|
| 67 |
+
]
|
| 68 |
+
system_prompt = "You are a helpful assistant."
|
| 69 |
+
|
| 70 |
+
[good_prompts]
|
| 71 |
+
dataset = "mlabonne/harmless_alpaca"
|
| 72 |
+
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
|
| 73 |
+
split = "train[:400]"
|
| 74 |
+
column = "text"
|
| 75 |
+
prefix = ""
|
| 76 |
+
suffix = ""
|
| 77 |
+
|
| 78 |
+
[bad_prompts]
|
| 79 |
+
dataset = "mlabonne/harmful_behaviors"
|
| 80 |
+
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
|
| 81 |
+
split = "train[:400]"
|
| 82 |
+
column = "text"
|
| 83 |
+
prefix = ""
|
| 84 |
+
suffix = ""
|
| 85 |
+
|
| 86 |
+
[good_evaluation_prompts]
|
| 87 |
+
dataset = "mlabonne/harmless_alpaca"
|
| 88 |
+
commit = "02c6a92cfcf11bb0c387334f8146d149d65b587f"
|
| 89 |
+
split = "test[:100]"
|
| 90 |
+
column = "text"
|
| 91 |
+
prefix = ""
|
| 92 |
+
suffix = ""
|
| 93 |
+
|
| 94 |
+
[bad_evaluation_prompts]
|
| 95 |
+
dataset = "mlabonne/harmful_behaviors"
|
| 96 |
+
commit = "01cead01398926d81f7c52bdb790ee8cf77ebba7"
|
| 97 |
+
split = "test[:100]"
|
| 98 |
+
column = "text"
|
| 99 |
+
prefix = ""
|
| 100 |
+
suffix = ""
|