apunkt/doc-to-lora
1
0
Fork 0
mirror of https://github.com/SakanaAI/doc-to-lora.git synced 2026-07-23 17:01:04 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

add pwc

Browse source
This commit is contained in:
51616 2025-01-05 16:37:04 +00:00
parent 46c9ddb891
commit 6a2e61ccc8
4 changed files with 356 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

42
configs/pwc.yaml Normal file
View file

@ -0,0 +1,42 @@
output_dir: "" # just a placeholder
bf16: true
model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
label_names: ["labels"]
# eval_on_start: True
# eval_strategy: "steps"
# eval_steps: 500
# save_strategy: "no"
# # save_steps: 500
# logging_strategy: "steps"
# logging_steps: 100
# use_liger_kernel: true
# remove_unused_columns: false
# needed to avoid OOM by compute the metrics batch by batch
# w/o this the trainer stores logits of all sample in memory...
# batch_eval_metrics: true
per_device_train_batch_size: 32
per_device_eval_batch_size: 32
max_val_samples_per_ds: 20
# optim: schedule_free_adamw
learning_rate: 0.0001
# lr_scheduler_type: "constant_with_warmup"
neftune_noise_alpha: 1
weight_decay: 0.1
warmup_ratio: 0.05
# LoRA
lora_r: 16
lora_dropout: 0.05
target_modules:
- down_proj
- up_proj
- gate_proj
# data
train_ds_names:
- sggetao/PwC
test_ds_names:
- sggetao/PwC

306
configs/pwc_and_ctx_numbers_256.yaml Normal file
View file

@ -0,0 +1,306 @@
output_dir: "" # just a placeholder
bf16: true
model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
label_names: ["labels"]
# eval_on_start: True
# eval_strategy: "steps"
# eval_steps: 500
# save_strategy: "no"
# # save_steps: 500
# logging_strategy: "steps"
# logging_steps: 100
# use_liger_kernel: true
# remove_unused_columns: false
# needed to avoid OOM by compute the metrics batch by batch
# w/o this the trainer stores logits of all sample in memory...
# batch_eval_metrics: true
per_device_train_batch_size: 32
per_device_eval_batch_size: 32
max_val_samples_per_ds: 20
# optim: schedule_free_adamw
learning_rate: 0.0001
# lr_scheduler_type: "constant_with_warmup"
neftune_noise_alpha: 1
weight_decay: 0.1
warmup_ratio: 0.05
# LoRA
lora_r: 16
lora_dropout: 0.05
target_modules:
- down_proj
- up_proj
- gate_proj
# data
train_ds_names:
- sggetao/PwC
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_129
- data/raw_datasets/context_numbers_130
- data/raw_datasets/context_numbers_131
- data/raw_datasets/context_numbers_132
- data/raw_datasets/context_numbers_133
- data/raw_datasets/context_numbers_134
- data/raw_datasets/context_numbers_135
- data/raw_datasets/context_numbers_136
- data/raw_datasets/context_numbers_137
- data/raw_datasets/context_numbers_138
- data/raw_datasets/context_numbers_139
- data/raw_datasets/context_numbers_140
- data/raw_datasets/context_numbers_141
- data/raw_datasets/context_numbers_142
- data/raw_datasets/context_numbers_143
- data/raw_datasets/context_numbers_144
- data/raw_datasets/context_numbers_145
- data/raw_datasets/context_numbers_146
- data/raw_datasets/context_numbers_147
- data/raw_datasets/context_numbers_148
- data/raw_datasets/context_numbers_149
- data/raw_datasets/context_numbers_150
- data/raw_datasets/context_numbers_151
- data/raw_datasets/context_numbers_152
- data/raw_datasets/context_numbers_153
- data/raw_datasets/context_numbers_154
- data/raw_datasets/context_numbers_155
- data/raw_datasets/context_numbers_156
- data/raw_datasets/context_numbers_157
- data/raw_datasets/context_numbers_158
- data/raw_datasets/context_numbers_159
- data/raw_datasets/context_numbers_160
- data/raw_datasets/context_numbers_161
- data/raw_datasets/context_numbers_162
- data/raw_datasets/context_numbers_163
- data/raw_datasets/context_numbers_164
- data/raw_datasets/context_numbers_165
- data/raw_datasets/context_numbers_166
- data/raw_datasets/context_numbers_167
- data/raw_datasets/context_numbers_168
- data/raw_datasets/context_numbers_169
- data/raw_datasets/context_numbers_170
- data/raw_datasets/context_numbers_171
- data/raw_datasets/context_numbers_172
- data/raw_datasets/context_numbers_173
- data/raw_datasets/context_numbers_174
- data/raw_datasets/context_numbers_175
- data/raw_datasets/context_numbers_176
- data/raw_datasets/context_numbers_177
- data/raw_datasets/context_numbers_178
- data/raw_datasets/context_numbers_179
- data/raw_datasets/context_numbers_180
- data/raw_datasets/context_numbers_181
- data/raw_datasets/context_numbers_182
- data/raw_datasets/context_numbers_183
- data/raw_datasets/context_numbers_184
- data/raw_datasets/context_numbers_185
- data/raw_datasets/context_numbers_186
- data/raw_datasets/context_numbers_187
- data/raw_datasets/context_numbers_188
- data/raw_datasets/context_numbers_189
- data/raw_datasets/context_numbers_190
- data/raw_datasets/context_numbers_191
- data/raw_datasets/context_numbers_192
- data/raw_datasets/context_numbers_193
- data/raw_datasets/context_numbers_194
- data/raw_datasets/context_numbers_195
- data/raw_datasets/context_numbers_196
- data/raw_datasets/context_numbers_197
- data/raw_datasets/context_numbers_198
- data/raw_datasets/context_numbers_199
- data/raw_datasets/context_numbers_200
- data/raw_datasets/context_numbers_201
- data/raw_datasets/context_numbers_202
- data/raw_datasets/context_numbers_203
- data/raw_datasets/context_numbers_204
- data/raw_datasets/context_numbers_205
- data/raw_datasets/context_numbers_206
- data/raw_datasets/context_numbers_207
- data/raw_datasets/context_numbers_208
- data/raw_datasets/context_numbers_209
- data/raw_datasets/context_numbers_210
- data/raw_datasets/context_numbers_211
- data/raw_datasets/context_numbers_212
- data/raw_datasets/context_numbers_213
- data/raw_datasets/context_numbers_214
- data/raw_datasets/context_numbers_215
- data/raw_datasets/context_numbers_216
- data/raw_datasets/context_numbers_217
- data/raw_datasets/context_numbers_218
- data/raw_datasets/context_numbers_219
- data/raw_datasets/context_numbers_220
- data/raw_datasets/context_numbers_221
- data/raw_datasets/context_numbers_222
- data/raw_datasets/context_numbers_223
- data/raw_datasets/context_numbers_224
- data/raw_datasets/context_numbers_225
- data/raw_datasets/context_numbers_226
- data/raw_datasets/context_numbers_227
- data/raw_datasets/context_numbers_228
- data/raw_datasets/context_numbers_229
- data/raw_datasets/context_numbers_230
- data/raw_datasets/context_numbers_231
- data/raw_datasets/context_numbers_232
- data/raw_datasets/context_numbers_233
- data/raw_datasets/context_numbers_234
- data/raw_datasets/context_numbers_235
- data/raw_datasets/context_numbers_236
- data/raw_datasets/context_numbers_237
- data/raw_datasets/context_numbers_238
- data/raw_datasets/context_numbers_239
- data/raw_datasets/context_numbers_240
- data/raw_datasets/context_numbers_241
- data/raw_datasets/context_numbers_242
- data/raw_datasets/context_numbers_243
- data/raw_datasets/context_numbers_244
- data/raw_datasets/context_numbers_245
- data/raw_datasets/context_numbers_246
- data/raw_datasets/context_numbers_247
- data/raw_datasets/context_numbers_248
- data/raw_datasets/context_numbers_249
- data/raw_datasets/context_numbers_250
- data/raw_datasets/context_numbers_251
- data/raw_datasets/context_numbers_252
- data/raw_datasets/context_numbers_253
- data/raw_datasets/context_numbers_254
- data/raw_datasets/context_numbers_255
- data/raw_datasets/context_numbers_256
val_ds_names:
- data/raw_datasets/context_numbers_128
- data/raw_datasets/context_numbers_129
- data/raw_datasets/context_numbers_130
- data/raw_datasets/context_numbers_131
- data/raw_datasets/context_numbers_132
- data/raw_datasets/context_numbers_133
- data/raw_datasets/context_numbers_134
- data/raw_datasets/context_numbers_135
- data/raw_datasets/context_numbers_136
- data/raw_datasets/context_numbers_137
- data/raw_datasets/context_numbers_138
- data/raw_datasets/context_numbers_139
- data/raw_datasets/context_numbers_140
- data/raw_datasets/context_numbers_141
- data/raw_datasets/context_numbers_142
- data/raw_datasets/context_numbers_143
- data/raw_datasets/context_numbers_144
- data/raw_datasets/context_numbers_145
- data/raw_datasets/context_numbers_146
- data/raw_datasets/context_numbers_147
- data/raw_datasets/context_numbers_148
- data/raw_datasets/context_numbers_149
- data/raw_datasets/context_numbers_150
- data/raw_datasets/context_numbers_151
- data/raw_datasets/context_numbers_152
- data/raw_datasets/context_numbers_153
- data/raw_datasets/context_numbers_154
- data/raw_datasets/context_numbers_155
- data/raw_datasets/context_numbers_156
- data/raw_datasets/context_numbers_157
- data/raw_datasets/context_numbers_158
- data/raw_datasets/context_numbers_159
- data/raw_datasets/context_numbers_160
- data/raw_datasets/context_numbers_161
- data/raw_datasets/context_numbers_162
- data/raw_datasets/context_numbers_163
- data/raw_datasets/context_numbers_164
- data/raw_datasets/context_numbers_165
- data/raw_datasets/context_numbers_166
- data/raw_datasets/context_numbers_167
- data/raw_datasets/context_numbers_168
- data/raw_datasets/context_numbers_169
- data/raw_datasets/context_numbers_170
- data/raw_datasets/context_numbers_171
- data/raw_datasets/context_numbers_172
- data/raw_datasets/context_numbers_173
- data/raw_datasets/context_numbers_174
- data/raw_datasets/context_numbers_175
- data/raw_datasets/context_numbers_176
- data/raw_datasets/context_numbers_177
- data/raw_datasets/context_numbers_178
- data/raw_datasets/context_numbers_179
- data/raw_datasets/context_numbers_180
- data/raw_datasets/context_numbers_181
- data/raw_datasets/context_numbers_182
- data/raw_datasets/context_numbers_183
- data/raw_datasets/context_numbers_184
- data/raw_datasets/context_numbers_185
- data/raw_datasets/context_numbers_186
- data/raw_datasets/context_numbers_187
- data/raw_datasets/context_numbers_188
- data/raw_datasets/context_numbers_189
- data/raw_datasets/context_numbers_190
- data/raw_datasets/context_numbers_191
- data/raw_datasets/context_numbers_192
- data/raw_datasets/context_numbers_193
- data/raw_datasets/context_numbers_194
- data/raw_datasets/context_numbers_195
- data/raw_datasets/context_numbers_196
- data/raw_datasets/context_numbers_197
- data/raw_datasets/context_numbers_198
- data/raw_datasets/context_numbers_199
- data/raw_datasets/context_numbers_200
- data/raw_datasets/context_numbers_201
- data/raw_datasets/context_numbers_202
- data/raw_datasets/context_numbers_203
- data/raw_datasets/context_numbers_204
- data/raw_datasets/context_numbers_205
- data/raw_datasets/context_numbers_206
- data/raw_datasets/context_numbers_207
- data/raw_datasets/context_numbers_208
- data/raw_datasets/context_numbers_209
- data/raw_datasets/context_numbers_210
- data/raw_datasets/context_numbers_211
- data/raw_datasets/context_numbers_212
- data/raw_datasets/context_numbers_213
- data/raw_datasets/context_numbers_214
- data/raw_datasets/context_numbers_215
- data/raw_datasets/context_numbers_216
- data/raw_datasets/context_numbers_217
- data/raw_datasets/context_numbers_218
- data/raw_datasets/context_numbers_219
- data/raw_datasets/context_numbers_220
- data/raw_datasets/context_numbers_221
- data/raw_datasets/context_numbers_222
- data/raw_datasets/context_numbers_223
- data/raw_datasets/context_numbers_224
- data/raw_datasets/context_numbers_225
- data/raw_datasets/context_numbers_226
- data/raw_datasets/context_numbers_227
- data/raw_datasets/context_numbers_228
- data/raw_datasets/context_numbers_229
- data/raw_datasets/context_numbers_230
- data/raw_datasets/context_numbers_231
- data/raw_datasets/context_numbers_232
- data/raw_datasets/context_numbers_233
- data/raw_datasets/context_numbers_234
- data/raw_datasets/context_numbers_235
- data/raw_datasets/context_numbers_236
- data/raw_datasets/context_numbers_237
- data/raw_datasets/context_numbers_238
- data/raw_datasets/context_numbers_239
- data/raw_datasets/context_numbers_240
- data/raw_datasets/context_numbers_241
- data/raw_datasets/context_numbers_242
- data/raw_datasets/context_numbers_243
- data/raw_datasets/context_numbers_244
- data/raw_datasets/context_numbers_245
- data/raw_datasets/context_numbers_246
- data/raw_datasets/context_numbers_247
- data/raw_datasets/context_numbers_248
- data/raw_datasets/context_numbers_249
- data/raw_datasets/context_numbers_250
- data/raw_datasets/context_numbers_251
- data/raw_datasets/context_numbers_252
- data/raw_datasets/context_numbers_253
- data/raw_datasets/context_numbers_254
- data/raw_datasets/context_numbers_255
- data/raw_datasets/context_numbers_256
test_ds_names:
- data/raw_datasets/context_numbers_512
- data/raw_datasets/context_numbers_1024
- data/raw_datasets/context_numbers_2048
- sggetao/PwC

7
hyperlora/data_utils.py
View file

@ -188,6 +188,13 @@ def get_preprocessing_fn(ds_name: str) -> Callable[[dict[str, Any]], dict[str, A
"""
f = lambda x: x
if ds_name == "sggetao/PwC":
def f(example):
example["context"] = example["input"]
example["response"] = example["answer"]
return example
# if ds_name == "context_numbers":
# def f(example):

2
hyperlora/intx_sft.py
View file

@ -276,7 +276,7 @@ def main(output_dir):
["train", "validation", "test"],
[data_args.train_ds_names, data_args.val_ds_names, data_args.test_ds_names],
):
if ds_names is None:
if not ds_names:
continue
if split == "train":
tokenized_ds[split] = concatenate_datasets(