mirror of
https://github.com/wassname/discovering_latent_knowledge.git
synced 2026-09-17 12:10:05 +08:00
misc
This commit is contained in:
wassname
+4
-1
@@ -1821,4 +1821,7 @@ Ok what should my pipeline do?
|
||||
|
||||
|
||||
TODO:
|
||||
- [ ] change layer -1 to actual name?
|
||||
- [x] change layer -1 to actual name?
|
||||
- [ ] in the TQS notebook they only get the mean from the generated tokens
|
||||
- [ ] they only interven on the ones that are not padded... wait no they passed in no token pos in honesty
|
||||
- [ ] but they did in TQA
|
||||
|
||||
+277
-83
@@ -87,12 +87,12 @@
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"\u001b[32m2023-10-24 10:49:20.938\u001b[0m | \u001b[1mINFO \u001b[0m | \u001b[36msrc.models.load\u001b[0m:\u001b[36mverbose_change_param\u001b[0m:\u001b[36m18\u001b[0m - \u001b[1mchanging pad_token_id from 32000 to 0\u001b[0m\n",
|
||||
"2023-10-24T10:49:20.938690+0800 INFO changing pad_token_id from 32000 to 0\n",
|
||||
"\u001b[32m2023-10-24 10:49:20.940\u001b[0m | \u001b[1mINFO \u001b[0m | \u001b[36msrc.models.load\u001b[0m:\u001b[36mverbose_change_param\u001b[0m:\u001b[36m18\u001b[0m - \u001b[1mchanging padding_side from right to left\u001b[0m\n",
|
||||
"2023-10-24T10:49:20.940322+0800 INFO changing padding_side from right to left\n",
|
||||
"\u001b[32m2023-10-24 10:49:20.941\u001b[0m | \u001b[1mINFO \u001b[0m | \u001b[36msrc.models.load\u001b[0m:\u001b[36mverbose_change_param\u001b[0m:\u001b[36m18\u001b[0m - \u001b[1mchanging truncation_side from right to left\u001b[0m\n",
|
||||
"2023-10-24T10:49:20.941383+0800 INFO changing truncation_side from right to left\n"
|
||||
"\u001b[32m2023-10-24 12:20:50.811\u001b[0m | \u001b[1mINFO \u001b[0m | \u001b[36msrc.models.load\u001b[0m:\u001b[36mverbose_change_param\u001b[0m:\u001b[36m18\u001b[0m - \u001b[1mchanging pad_token_id from 32000 to 0\u001b[0m\n",
|
||||
"2023-10-24T12:20:50.811435+0800 INFO changing pad_token_id from 32000 to 0\n",
|
||||
"\u001b[32m2023-10-24 12:20:50.812\u001b[0m | \u001b[1mINFO \u001b[0m | \u001b[36msrc.models.load\u001b[0m:\u001b[36mverbose_change_param\u001b[0m:\u001b[36m18\u001b[0m - \u001b[1mchanging padding_side from right to left\u001b[0m\n",
|
||||
"2023-10-24T12:20:50.812537+0800 INFO changing padding_side from right to left\n",
|
||||
"\u001b[32m2023-10-24 12:20:50.813\u001b[0m | \u001b[1mINFO \u001b[0m | \u001b[36msrc.models.load\u001b[0m:\u001b[36mverbose_change_param\u001b[0m:\u001b[36m18\u001b[0m - \u001b[1mchanging truncation_side from right to left\u001b[0m\n",
|
||||
"2023-10-24T12:20:50.813021+0800 INFO changing truncation_side from right to left\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
@@ -147,11 +147,11 @@
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"tokenize: 100%|██████████| 302/302 [00:00<00:00, 3490.12 examples/s]\n",
|
||||
"truncated: 100%|██████████| 302/302 [00:00<00:00, 3889.90 examples/s]\n",
|
||||
"prompt_truncated: 100%|██████████| 302/302 [00:00<00:00, 604.08 examples/s]\n",
|
||||
"choice_ids: 100%|██████████| 302/302 [00:00<00:00, 9749.77 examples/s]\n",
|
||||
"Filter: 100%|██████████| 302/302 [00:00<00:00, 3987.01 examples/s]"
|
||||
"tokenize: 100%|██████████| 302/302 [00:00<00:00, 2898.85 examples/s]\n",
|
||||
"truncated: 100%|██████████| 302/302 [00:00<00:00, 3922.12 examples/s]\n",
|
||||
"prompt_truncated: 100%|██████████| 302/302 [00:00<00:00, 565.78 examples/s]\n",
|
||||
"choice_ids: 100%|██████████| 302/302 [00:00<00:00, 10226.05 examples/s]\n",
|
||||
"Filter: 100%|██████████| 302/302 [00:00<00:00, 3935.72 examples/s]"
|
||||
]
|
||||
},
|
||||
{
|
||||
@@ -236,7 +236,7 @@
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<src.repe.rep_readers.PCARepReader at 0x7fcd97b27250>"
|
||||
"<src.repe.rep_readers.PCARepReader at 0x7ff755ec4ee0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
@@ -268,17 +268,17 @@
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"{8: -7.1495447,\n",
|
||||
" 11: 16.796896,\n",
|
||||
" 14: 13.5725565,\n",
|
||||
" 17: -5.684828,\n",
|
||||
" 20: 26.571663,\n",
|
||||
" 23: 37.01544,\n",
|
||||
" 26: -67.76076,\n",
|
||||
" 29: -64.502205,\n",
|
||||
" 32: -52.72602,\n",
|
||||
" 35: -76.70331,\n",
|
||||
" 38: -80.74953}"
|
||||
"{8: 1.661585,\n",
|
||||
" 11: 4.2028227,\n",
|
||||
" 14: -1.7875421,\n",
|
||||
" 17: -6.6559744,\n",
|
||||
" 20: -7.8236346,\n",
|
||||
" 23: -10.668092,\n",
|
||||
" 26: -1.1562811,\n",
|
||||
" 29: -2.477594,\n",
|
||||
" 32: -3.406129,\n",
|
||||
" 35: -6.056723,\n",
|
||||
" 38: 3.0787961}"
|
||||
]
|
||||
},
|
||||
"execution_count": 9,
|
||||
@@ -308,61 +308,241 @@
|
||||
"cell_type": "code",
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<src.repe.rep_control_pipeline.RepControlPipeline at 0x7ff755ec4ca0>"
|
||||
]
|
||||
},
|
||||
"execution_count": 10,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"\n",
|
||||
"# layer_id = hidden_layers\n",
|
||||
"# block_name=\"decoder_block\"\n",
|
||||
"# control_method=\"reading_vec\"\n",
|
||||
"layer_id = hidden_layers\n",
|
||||
"block_name=\"decoder_block\"\n",
|
||||
"control_method=\"reading_vec\"\n",
|
||||
"\n",
|
||||
"# rep_control_pipeline = pipeline(\n",
|
||||
"# \"rep-control\", \n",
|
||||
"# model=model, \n",
|
||||
"# tokenizer=tokenizer, \n",
|
||||
"# layers=layer_id, max_length=cfg.max_length,\n",
|
||||
"# control_method=control_method)\n",
|
||||
"# rep_control_pipeline\n"
|
||||
"rep_control_pipeline = pipeline(\n",
|
||||
" \"rep-control\", \n",
|
||||
" model=model, \n",
|
||||
" tokenizer=tokenizer, \n",
|
||||
" layers=layer_id, max_length=cfg.max_length,\n",
|
||||
" control_method=control_method)\n",
|
||||
"rep_control_pipeline\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 11,
|
||||
"execution_count": 16,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"===== No Control =====\n",
|
||||
"negative\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"negative\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"I\n",
|
||||
"\n",
|
||||
"===== No Control =====\n",
|
||||
"pos\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"pos\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"pos\n",
|
||||
"\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"\n",
|
||||
"# from re import S\n",
|
||||
"from re import S\n",
|
||||
"layer_id = hidden_layers\n",
|
||||
"\n",
|
||||
"inputs = dataset_train[:2]\n",
|
||||
"coeff=8.0\n",
|
||||
"max_new_tokens=1\n",
|
||||
"text_gen_kwargs = dict(do_sample=False, max_new_tokens=max_new_tokens, use_cache=False, \n",
|
||||
" output_hidden_states=True, return_dict=True,\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"# inputs = dataset_train[:2]\n",
|
||||
"# coeff=8.0\n",
|
||||
"# max_new_tokens=1\n",
|
||||
"# text_gen_kwargs = dict(do_sample=False, max_new_tokens=max_new_tokens, use_cache=False, \n",
|
||||
"# output_hidden_states=True, return_dict=True,\n",
|
||||
"# )\n",
|
||||
"\n",
|
||||
"# activations = {}\n",
|
||||
"# for layer in layer_id:\n",
|
||||
"# activations[layer] = torch.tensor(coeff * honesty_rep_reader.directions[layer] * honesty_rep_reader.direction_signs[layer]).to(model.device).half()\n",
|
||||
"activations = {}\n",
|
||||
"for layer in layer_id:\n",
|
||||
" activations[layer] = torch.tensor(coeff * honesty_rep_reader.directions[layer] * honesty_rep_reader.direction_signs[layer]).to(model.device).half()\n",
|
||||
" \n",
|
||||
"\n",
|
||||
"# activations_neg = {k:-v for k,v in activations.items()}\n",
|
||||
"activations_neg = {k:-v for k,v in activations.items()}\n",
|
||||
"\n",
|
||||
"# model.eval()\n",
|
||||
"# with torch.no_grad():\n",
|
||||
"# baseline_outputs = rep_control_pipeline(inputs, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
"# control_outputs = rep_control_pipeline(inputs, activations=activations, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
"# control_outputs_neg = rep_control_pipeline(inputs, activations=activations_neg, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
"model.eval()\n",
|
||||
"with torch.no_grad():\n",
|
||||
" baseline_outputs = rep_control_pipeline(inputs, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
" control_outputs = rep_control_pipeline(inputs, activations=activations, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
" control_outputs_neg = rep_control_pipeline(inputs, activations=activations_neg, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
"\n",
|
||||
"# for i,s,p,n in zip(inputs, baseline_outputs['text_ans'], control_outputs['text_ans'], control_outputs_neg['text_ans']):\n",
|
||||
"# print(\"===== No Control =====\")\n",
|
||||
"# print(s)\n",
|
||||
"# print(f\"===== + Honesty Control =====\")\n",
|
||||
"# print(p)\n",
|
||||
"# print()\n",
|
||||
"# print(f\"===== - Honesty Control =====\")\n",
|
||||
"# print(n)\n",
|
||||
"# print()\n"
|
||||
"for i,s,p,n in zip(inputs, baseline_outputs['text_ans'], control_outputs['text_ans'], control_outputs_neg['text_ans']):\n",
|
||||
" print(\"===== No Control =====\")\n",
|
||||
" print(s)\n",
|
||||
" print(f\"===== + Honesty Control =====\")\n",
|
||||
" print(p)\n",
|
||||
" print()\n",
|
||||
" print(f\"===== - Honesty Control =====\")\n",
|
||||
" print(n)\n",
|
||||
" print()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 17,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"===== No Control =====\n",
|
||||
"15.20%\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"19.81%\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"29.42%\n",
|
||||
"\n",
|
||||
"===== No Control =====\n",
|
||||
"61.88%\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"91.96%\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"84.79%\n",
|
||||
"\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"for i,s,p,n in zip(inputs, baseline_outputs['ans'], control_outputs['ans'], control_outputs_neg['ans']):\n",
|
||||
" print(\"===== No Control =====\")\n",
|
||||
" print(f'{s:02.2%}')\n",
|
||||
" print(f\"===== + Honesty Control =====\")\n",
|
||||
" print(f'{p:02.2%}')\n",
|
||||
" print()\n",
|
||||
" print(f\"===== - Honesty Control =====\")\n",
|
||||
" print(f'{n:02.2%}')\n",
|
||||
" print()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 21,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"===== No Control =====\n",
|
||||
"negative\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"negative\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"I\n",
|
||||
"\n",
|
||||
"===== No Control =====\n",
|
||||
"pos\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"pos\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"pos\n",
|
||||
"\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"\n",
|
||||
"from re import S\n",
|
||||
"layer_id = hidden_layers\n",
|
||||
"\n",
|
||||
"inputs = dataset_train[:2]\n",
|
||||
"coeff=8.0\n",
|
||||
"max_new_tokens=1\n",
|
||||
"text_gen_kwargs = dict(do_sample=False, max_new_tokens=max_new_tokens, use_cache=False, \n",
|
||||
" output_hidden_states=True, return_dict=True,\n",
|
||||
" token_pos=\"end\",\n",
|
||||
" normalize=False\n",
|
||||
" )\n",
|
||||
"\n",
|
||||
"activations = {}\n",
|
||||
"for layer in layer_id:\n",
|
||||
" activations[layer] = torch.tensor(coeff * honesty_rep_reader.directions[layer] * honesty_rep_reader.direction_signs[layer]).to(model.device).half()\n",
|
||||
" \n",
|
||||
"\n",
|
||||
"activations_neg = {k:-v for k,v in activations.items()}\n",
|
||||
"\n",
|
||||
"model.eval()\n",
|
||||
"with torch.no_grad():\n",
|
||||
" baseline_outputs = rep_control_pipeline(inputs, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
" control_outputs = rep_control_pipeline(inputs, activations=activations, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
" control_outputs_neg = rep_control_pipeline(inputs, activations=activations_neg, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
"\n",
|
||||
"for i,s,p,n in zip(inputs, baseline_outputs['text_ans'], control_outputs['text_ans'], control_outputs_neg['text_ans']):\n",
|
||||
" print(\"===== No Control =====\")\n",
|
||||
" print(s)\n",
|
||||
" print(f\"===== + Honesty Control =====\")\n",
|
||||
" print(p)\n",
|
||||
" print()\n",
|
||||
" print(f\"===== - Honesty Control =====\")\n",
|
||||
" print(n)\n",
|
||||
" print()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 22,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"===== No Control =====\n",
|
||||
"15.20%\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"19.81%\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"29.42%\n",
|
||||
"\n",
|
||||
"===== No Control =====\n",
|
||||
"61.88%\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"91.96%\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"84.79%\n",
|
||||
"\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"for i,s,p,n in zip(inputs, baseline_outputs['ans'], control_outputs['ans'], control_outputs_neg['ans']):\n",
|
||||
" print(\"===== No Control =====\")\n",
|
||||
" print(f'{s:02.2%}')\n",
|
||||
" print(f\"===== + Honesty Control =====\")\n",
|
||||
" print(f'{p:02.2%}')\n",
|
||||
" print()\n",
|
||||
" print(f\"===== - Honesty Control =====\")\n",
|
||||
" print(f'{n:02.2%}')\n",
|
||||
" print()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
@@ -381,35 +561,35 @@
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 12,
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<src.repe.rep_control_pipeline_baukit.RepControlPipeline2 at 0x7fcca5807130>"
|
||||
"<src.repe.rep_control_pipeline_baukit.RepControlPipeline2 at 0x7ff755f5be50>"
|
||||
]
|
||||
},
|
||||
"execution_count": 12,
|
||||
"execution_count": 13,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"layer_id = hidden_layers\n",
|
||||
"\n",
|
||||
"rep_control_pipeline = pipeline(\n",
|
||||
"\n",
|
||||
"rep_control_pipeline2 = pipeline(\n",
|
||||
" \"rep-control2\", \n",
|
||||
" model=model, \n",
|
||||
" tokenizer=tokenizer, \n",
|
||||
" layers=layer_id, \n",
|
||||
" max_length=cfg.max_length,)\n",
|
||||
"rep_control_pipeline\n"
|
||||
"rep_control_pipeline2\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 47,
|
||||
"execution_count": 14,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
@@ -422,15 +602,15 @@
|
||||
"negative\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"negative\n",
|
||||
"I\n",
|
||||
"\n",
|
||||
"===== No Control =====\n",
|
||||
"The\n",
|
||||
"pos\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"The\n",
|
||||
"pos\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"negative\n",
|
||||
"pos\n",
|
||||
"\n"
|
||||
]
|
||||
}
|
||||
@@ -452,9 +632,9 @@
|
||||
"\n",
|
||||
"model.eval()\n",
|
||||
"with torch.no_grad():\n",
|
||||
" baseline_outputs = rep_control_pipeline(inputs, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
" control_outputs = rep_control_pipeline(inputs, activations=activations, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
" control_outputs_neg = rep_control_pipeline(inputs, activations=activations_neg, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
" baseline_outputs = rep_control_pipeline2(inputs, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
" control_outputs = rep_control_pipeline2(inputs, activations=activations, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
" control_outputs_neg = rep_control_pipeline2(inputs, activations=activations_neg, batch_size=batch_size, **text_gen_kwargs)\n",
|
||||
"\n",
|
||||
"\n",
|
||||
"for i,s,p,n in zip(inputs, baseline_outputs['text_ans'], control_outputs['text_ans'], control_outputs_neg['text_ans']):\n",
|
||||
@@ -470,7 +650,7 @@
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 71,
|
||||
"execution_count": 15,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
@@ -478,20 +658,20 @@
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"===== No Control =====\n",
|
||||
"2.03%\n",
|
||||
"15.20%\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"23.10%\n",
|
||||
"19.81%\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"1.02%\n",
|
||||
"29.42%\n",
|
||||
"\n",
|
||||
"===== No Control =====\n",
|
||||
"33.63%\n",
|
||||
"61.88%\n",
|
||||
"===== + Honesty Control =====\n",
|
||||
"83.52%\n",
|
||||
"91.96%\n",
|
||||
"\n",
|
||||
"===== - Honesty Control =====\n",
|
||||
"7.75%\n",
|
||||
"84.79%\n",
|
||||
"\n"
|
||||
]
|
||||
}
|
||||
@@ -508,6 +688,20 @@
|
||||
" print()\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
|
||||
@@ -35,7 +35,7 @@ class RepControlPipeline(RepControlPipeline2):
|
||||
self.layers = layers
|
||||
self.max_length = max_length
|
||||
|
||||
super().__init__(model=model, tokenizer=tokenizer, **kwargs)
|
||||
super().__init__(model=model, tokenizer=tokenizer, max_length=max_length, **kwargs)
|
||||
|
||||
def __call__(self, text_inputs, activations=None, **kwargs):
|
||||
if activations is not None:
|
||||
|
||||
@@ -20,6 +20,7 @@ def row_choice_ids(answer_choices, tokenizer):
|
||||
|
||||
|
||||
def intervene(output, activation):
|
||||
# TODO need attention mask
|
||||
assert output.ndim == 3, f"expected output to be (batch, seq, vocab), got {output.shape}"
|
||||
return output + activation.to(output.device)[None, None, :]
|
||||
|
||||
|
||||
Reference in New Issue
Block a user