Sentovara commited on
Commit
4d54b76
·
verified ·
1 Parent(s): c41d230

Sentinel S1 release: adapter, model card, evaluation code, results

Browse files
README.md ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: allenai/Olmo-3.1-32B-Instruct
4
+ library_name: peft
5
+ language:
6
+ - en
7
+ pipeline_tag: text-generation
8
+ tags:
9
+ - lora
10
+ - olmo
11
+ - uncertainty
12
+ - calibration
13
+ - abstention
14
+ - retrieval-augmented-generation
15
+ ---
16
+
17
+ # Sentinel S1 (32B)
18
+
19
+ **An open model that tells you when it doesn't know.**
20
+
21
+ Sentinel S1 32B is a LoRA adapter for [allenai/Olmo-3.1-32B-Instruct](https://huggingface.co/allenai/Olmo-3.1-32B-Instruct) (revision `ac0587e4a7744a551c059d8cd17ba220bc940dae`), built by [Sentovara](https://huggingface.co/Sentovara). When the model is likely to be wrong it starts its reply with **`UNSURE:`**. A system around it can then look the answer up instead of trusting a guess. When it answers plainly, the answer is far more likely to be right than the original model's.
22
+
23
+ Full method, results and limits: [the write-up](https://docs.google.com/document/d/1elcP-xWo-c0brn-lTzhtKyTaBuGjf0fL5-wCRz5AfaM/edit?usp=sharing).
24
+
25
+ ## Results
26
+
27
+ All checks were scored by script, on question sets kept out of training. The original model and Sentinel ran through the same code. The pass marks were written down before any results existed.
28
+
29
+ | Check | Original 32B | Sentinel S1 32B | Pass mark | Result |
30
+ |:--|:--|:--|:--|:--|
31
+ | Instruction-following (IFEval, 541 prompts) | 85.6% | 85.0% | within 2 points | Pass |
32
+ | Maths (GSM8K test, 1,319 problems) | 94.7% | 94.7% | within 2 points | Pass |
33
+ | Plain (unflagged) answers right (9,255 questions) | 28.3% | **68.2%** | +10 points | Pass |
34
+ | Own wrong answers flagged | none | **87.5%** | at least half | Pass |
35
+ | Replies end cleanly | 99.6% | 99.6% | at least 95% | Pass |
36
+ | Flagged questions right after searching (6,448) | 10.9% as a guess | **45.2%** | search beats guess | Pass |
37
+ | Plain answers right in conversation (350 fresh chats) | 55.0% | **77.9%** | +10 points | Pass |
38
+ | Own mistakes flagged in conversation | none | **78.8%** | at least half | Pass |
39
+ | Ordinary chats flagged | — | 3 of 50 | at most 10% | Pass |
40
+
41
+ 9 of 9 checks passed. The sign-off questions came from Natural Questions (validation), SimpleQA, GSM8K (test) and IFEval.
42
+
43
+ **What changed and what didn't.** Sentinel learned nothing new about the world. Across all test questions it gets about the same share right as the original. The gain is in judgement: it marks the answers it is likely to get wrong, so its plain answers can be trusted far more, and its flags show exactly when to look something up.
44
+
45
+ **Against the 7B Sentinel.** The recipe was run again unchanged on the 32B, to test whether it carries over to a bigger model. It does. The 32B starts out knowing more: 32.5% of general-knowledge tries were right before training, against 14.9% for the 7B. So it answers plainly more often and flags a slightly smaller share of its mistakes (87.5% against 94.0%). It is the stronger of the two at instructions, maths and answering after a search. The 7B version is at [Sentovara/Sentinel-S1-7B](https://huggingface.co/Sentovara/Sentinel-S1-7B).
46
+
47
+ ## How it was trained
48
+
49
+ 1. **Evidence.** The original model answered 10,000 questions with known answers (5,000 general knowledge from Natural Questions, 5,000 maths from GSM8K), five tries each. Every answer was marked right or wrong by script.
50
+ 2. **Labels from its own record.** Where the model was right on 4 or 5 of 5 tries, it was trained to give its own answer plainly. Where it was right on 0 or 1 of 5, it was trained to give the same answer starting with `UNSURE:`. Every answer it learned from on those questions was its own.
51
+ 3. **Keeping its skills.** Training also included general conversation practice and everyday chat, so Sentinel kept its existing skills and answers small talk normally.
52
+ The evaluation code is in `eval/` and the result summaries behind the tables above are in `results/`.
53
+
54
+ ## How to use it
55
+
56
+ ```python
57
+ from transformers import AutoModelForCausalLM, AutoTokenizer
58
+ from peft import PeftModel
59
+
60
+ base, rev = "allenai/Olmo-3.1-32B-Instruct", "ac0587e4a7744a551c059d8cd17ba220bc940dae"
61
+ tok = AutoTokenizer.from_pretrained(base, revision=rev)
62
+ model = AutoModelForCausalLM.from_pretrained(base, revision=rev, dtype="bfloat16", device_map="auto")
63
+ model = PeftModel.from_pretrained(model, "Sentovara/Sentinel-S1-32B")
64
+
65
+ msgs = [{"role": "user", "content": "Who composed the music for the 1968 film 2001: A Space Odyssey?"}]
66
+ ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
67
+ out = model.generate(ids, max_new_tokens=128, do_sample=False)
68
+ reply = tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True)
69
+ if reply.startswith("UNSURE:"):
70
+ print("Sentinel is unsure. Look this up before trusting it:", reply)
71
+ else:
72
+ print(reply)
73
+ ```
74
+
75
+ The adapter also works with vLLM's LoRA support (`enable_lora=True`, `max_lora_rank=16`). Tested with transformers 5.22.1, peft 0.21.2 and vLLM 0.31.0.
76
+
77
+ **Search step.** For a flagged question, retrieve Wikipedia passages, number them, and ask Sentinel with this prompt, followed by the numbered passages and then `Question: ...`:
78
+
79
+ > Answer the question using the Wikipedia passages below. Give a short factual answer only (a few words), then on a new line write 'Source: ' followed by the number of the passage you used.
80
+
81
+ Our search took the top 50 BM25 passages from each of two Wikipedia indexes (the 2018 DPR index and a November 2023 dump), reranked them with BAAI/bge-reranker-v2-m3, and passed the top 10 to Sentinel. The code is in `eval/`.
82
+
83
+ ## Limits
84
+
85
+ - **Search helps, but more than half of flagged questions still don't end in a right answer.** The system names a checked source for about three in four answers.
86
+ - **Maths flagging is weak.** Sentinel flags only a small share of its own maths mistakes. Most maths errors are still stated plainly.
87
+ - **The marking is strict.** Answers are marked right only if they contain the expected answer, so some correct answers phrased differently are marked wrong.
88
+ - **Tested on public English question sets and everyday chat only.** It has not been tested on other kinds of work or other languages. Do not rely on it for medical, legal, financial or safety-critical decisions.
89
+ - **One model family.** The recipe works on the 7B and 32B Olmo models. It has not been tested on other model families.
90
+
91
+ ## Credits and licences
92
+
93
+ - **Base model:** allenai/Olmo-3.1-32B-Instruct by Ai2 (Apache 2.0). This adapter is released under Apache 2.0.
94
+ - **Training and test data:** Natural Questions (CC BY-SA 3.0), GSM8K (MIT), IFEval (Apache 2.0; its checking code is included in `eval/ifeval/`), SimpleQA (MIT), and Ai2's Dolci instruction data (ODC-BY).
95
+ - **Search:** the Pyserini/Anserini Wikipedia index (University of Waterloo), English Wikipedia November 2023 from the Wikimedia Foundation (CC BY-SA 3.0 / GFDL), and BAAI/bge-reranker-v2-m3 (Apache 2.0).
96
+ - **Compute:** trained and tested on AWS. About $97 of GPU time.
97
+
98
+ Built by Sentovara. Designed, run and documented by Claude (Anthropic) for Sentovara.
adapter_config.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "allenai/Olmo-3.1-32B-Instruct",
7
+ "bias": "none",
8
+ "corda_config": null,
9
+ "ensure_weight_tying": false,
10
+ "eva_config": null,
11
+ "exclude_modules": null,
12
+ "fan_in_fan_out": false,
13
+ "inference_mode": true,
14
+ "init_lora_weights": true,
15
+ "kasa_config": null,
16
+ "layer_replication": null,
17
+ "layers_pattern": null,
18
+ "layers_to_transform": null,
19
+ "loftq_config": {},
20
+ "lora_alpha": 32,
21
+ "lora_bias": false,
22
+ "lora_dropout": 0.05,
23
+ "lora_ga_config": null,
24
+ "megatron_config": null,
25
+ "megatron_core": "megatron.core",
26
+ "modules_to_save": null,
27
+ "monteclora_config": null,
28
+ "peft_type": "LORA",
29
+ "peft_version": "0.21.2",
30
+ "qalora_group_size": 16,
31
+ "r": 16,
32
+ "rank_pattern": {},
33
+ "revision": null,
34
+ "target_modules": [
35
+ "down_proj",
36
+ "v_proj",
37
+ "k_proj",
38
+ "gate_proj",
39
+ "o_proj",
40
+ "q_proj",
41
+ "up_proj"
42
+ ],
43
+ "target_parameters": null,
44
+ "task_type": "CAUSAL_LM",
45
+ "trainable_token_indices": null,
46
+ "use_bdlora": null,
47
+ "use_dora": false,
48
+ "use_qalora": false,
49
+ "use_rslora": false,
50
+ "velora_config": null
51
+ }
adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:546a6494b46670c2bed2a2a30bd15a87d2ed691da85c57dddc6d0bee2ccdce11
3
+ size 536991984
eval/ifeval/__init__.py ADDED
File without changes
eval/ifeval/instructions.py ADDED
@@ -0,0 +1,1619 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 The Google Research Authors.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ """Library of instructions.
16
+
17
+ Note: the following checker classes are defined here but are NOT registered in
18
+ `instructions_registry.INSTRUCTION_DICT` and are not referenced by any row in
19
+ the published `google/IFEval` dataset — they are dead code retained from the
20
+ upstream Google import (see TODOs in `instructions_registry.py`):
21
+
22
+ - `ConstrainedStartChecker`
23
+ - `RephraseChecker` (requires an `original_message` kwarg the dataset does not provide)
24
+ - `KeySentenceChecker`
25
+ - `RephraseParagraph`
26
+ """
27
+
28
+ import collections
29
+ import json
30
+ import logging
31
+ import random
32
+ import re
33
+ import string
34
+ from collections.abc import Sequence
35
+
36
+ import langdetect
37
+
38
+ from . import instructions_util
39
+
40
+
41
+ logger = logging.getLogger(__name__)
42
+
43
+ _InstructionArgsDtype = dict[str, int | str | Sequence[str]] | None
44
+
45
+ _LANGUAGES = instructions_util.LANGUAGE_CODES
46
+
47
+ # The relational operation for comparison.
48
+ _COMPARISON_RELATION = ("less than", "at least")
49
+
50
+ # The maximum number of sentences.
51
+ _MAX_NUM_SENTENCES = 20
52
+
53
+ # The number of placeholders.
54
+ _NUM_PLACEHOLDERS = 4
55
+
56
+ # The number of bullet lists.
57
+ _NUM_BULLETS = 5
58
+
59
+ # The options of constrained response.
60
+ _CONSTRAINED_RESPONSE_OPTIONS = (
61
+ "My answer is yes.",
62
+ "My answer is no.",
63
+ "My answer is maybe.",
64
+ )
65
+
66
+ # The options of starter keywords.
67
+ _STARTER_OPTIONS = (
68
+ "I would say",
69
+ "My answer is",
70
+ "I believe",
71
+ "In my opinion",
72
+ "I think",
73
+ "I reckon",
74
+ "I feel",
75
+ "From my perspective",
76
+ "As I see it",
77
+ "According to me",
78
+ "As far as I'm concerned",
79
+ "To my understanding",
80
+ "In my view",
81
+ "My take on it is",
82
+ "As per my perception",
83
+ )
84
+
85
+ # The options of ending keywords.
86
+ # TODO(jeffreyzhou) add more ending options
87
+ _ENDING_OPTIONS = ("Any other questions?", "Is there anything else I can help with?")
88
+
89
+ # The number of highlighted sections.
90
+ _NUM_HIGHLIGHTED_SECTIONS = 4
91
+
92
+ # The section splitter.
93
+ _SECTION_SPLITER = ("Section", "SECTION")
94
+
95
+ # The number of sections.
96
+ _NUM_SECTIONS = 5
97
+
98
+ # The number of paragraphs.
99
+ _NUM_PARAGRAPHS = 5
100
+
101
+ # The postscript marker.
102
+ _POSTSCRIPT_MARKER = ("P.S.", "P.P.S")
103
+
104
+ # The number of keywords.
105
+ _NUM_KEYWORDS = 2
106
+
107
+ # The occurrences of a single keyword.
108
+ _KEYWORD_FREQUENCY = 3
109
+
110
+ # The occurrences of a single letter.
111
+ _LETTER_FREQUENCY = 10
112
+
113
+ # The occurrences of words with all capital letters.
114
+ _ALL_CAPITAL_WORD_FREQUENCY = 20
115
+
116
+ # The number of words in the response.
117
+ _NUM_WORDS_LOWER_LIMIT = 100
118
+ _NUM_WORDS_UPPER_LIMIT = 500
119
+
120
+
121
+ class Instruction:
122
+ """An instruction template."""
123
+
124
+ def __init__(self, instruction_id):
125
+ self.id = instruction_id
126
+
127
+ def build_description(self, **kwargs):
128
+ raise NotImplementedError("`build_description` not implemented.")
129
+
130
+ def get_instruction_args(self):
131
+ raise NotImplementedError("`get_instruction_args` not implemented.")
132
+
133
+ def get_instruction_args_keys(self):
134
+ raise NotImplementedError("`get_instruction_args_keys` not implemented.")
135
+
136
+ def check_following(self, value):
137
+ raise NotImplementedError("`check_following` not implemented.")
138
+
139
+
140
+ class ResponseLanguageChecker(Instruction):
141
+ """Check the language of the entire response."""
142
+
143
+ def build_description(self, *, language=None):
144
+ """Build the instruction description.
145
+
146
+ Args:
147
+ language: A string representing the expected language of the response. The
148
+ language has to comply to the 97 types defined in
149
+ `langid.py` (https://pypi.org/project/langid/1.1.5/), which follows
150
+ ISO 639-1 codes (https://en.wikipedia.org/wiki/List_of_ISO_639-1_codes);
151
+ for example, `en` for English, `zh` for Chinese, `fr` for French.
152
+
153
+ Returns:
154
+ A string representing the instruction description.
155
+ """
156
+ self._language = language
157
+ if self._language is None:
158
+ self._language = random.choice(list(_LANGUAGES.keys()))
159
+ # TODO(tianjianlu): opens the description generation to more choices.
160
+ self._description_pattern = (
161
+ "Your ENTIRE response should be in {language} language, no other "
162
+ + "language is allowed."
163
+ )
164
+ return self._description_pattern.format(language=_LANGUAGES[self._language])
165
+
166
+ def get_instruction_args(self):
167
+ """Returns the keyword args of `build_description`."""
168
+ return {"language": self._language}
169
+
170
+ def get_instruction_args_keys(self):
171
+ """Returns the args keys of `build_description`."""
172
+ return ["language"]
173
+
174
+ def check_following(self, value):
175
+ """Check if the language of the entire response follows the instruction.
176
+
177
+ Args:
178
+ value: A string representing the response.
179
+
180
+ Returns:
181
+ True if the language of `value` follows instruction; otherwise False.
182
+ """
183
+ assert isinstance(value, str)
184
+
185
+ try:
186
+ return langdetect.detect(value) == self._language
187
+ except langdetect.LangDetectException as e:
188
+ # Count as instruction is followed.
189
+ logger.error(
190
+ "Unable to detect language for text %s due to %s", value, e
191
+ ) # refex: disable=pytotw.037
192
+ return True
193
+
194
+
195
+ class NumberOfSentences(Instruction):
196
+ """Check the number of sentences."""
197
+
198
+ def build_description(self, *, num_sentences=None, relation=None):
199
+ """Build the instruction description.
200
+
201
+ Args:
202
+ num_sentences: An integer specifying the number of sentences as a
203
+ threshold.
204
+ relation: A string in (`less than`, `at least`), defining the relational
205
+ operator for comparison.
206
+ Two relational comparisons are supported for now:
207
+ if 'less than', the actual number of sentences < the threshold;
208
+ if 'at least', the actual number of sentences >= the threshold.
209
+
210
+ Returns:
211
+ A string representing the instruction description.
212
+ """
213
+ # The number of sentences as a threshold for comparison.
214
+ self._num_sentences_threshold = num_sentences
215
+ if self._num_sentences_threshold is None or self._num_sentences_threshold < 0:
216
+ self._num_sentences_threshold = random.randint(1, _MAX_NUM_SENTENCES)
217
+
218
+ if relation is None:
219
+ self._comparison_relation = random.choice(_COMPARISON_RELATION)
220
+ elif relation not in _COMPARISON_RELATION:
221
+ raise ValueError(
222
+ "The supported relation for comparison must be in "
223
+ f"{_COMPARISON_RELATION}, but {relation} is given."
224
+ )
225
+ else:
226
+ self._comparison_relation = relation
227
+
228
+ self._description_pattern = (
229
+ "Your response should contain {relation} {num_sentences} sentences."
230
+ )
231
+ return self._description_pattern.format(
232
+ relation=self._comparison_relation,
233
+ num_sentences=self._num_sentences_threshold,
234
+ )
235
+
236
+ def get_instruction_args(self):
237
+ """Returns the keyword args of `build_description`."""
238
+ return {
239
+ "num_sentences": self._num_sentences_threshold,
240
+ "relation": self._comparison_relation,
241
+ }
242
+
243
+ def get_instruction_args_keys(self):
244
+ """Returns the args keys of `build_description`."""
245
+ return ["num_sentences", "relation"]
246
+
247
+ def check_following(self, value):
248
+ """Check if the number of sentences follows the instruction.
249
+
250
+ Args:
251
+ value: A string representing the response.
252
+
253
+ Returns:
254
+ True if the response follows the instruction.
255
+
256
+ Raise:
257
+ ValueError if the string in `instruction_args` is not in
258
+ [`less_than`, `at_least`].
259
+ """
260
+ num_sentences = instructions_util.count_sentences(value)
261
+ if self._comparison_relation == _COMPARISON_RELATION[0]:
262
+ return num_sentences < self._num_sentences_threshold
263
+ elif self._comparison_relation == _COMPARISON_RELATION[1]:
264
+ return num_sentences >= self._num_sentences_threshold
265
+
266
+
267
+ class PlaceholderChecker(Instruction):
268
+ """Check the placeholders in template writing."""
269
+
270
+ def build_description(self, *, num_placeholders=None):
271
+ """Build the instruction description.
272
+
273
+ Args:
274
+ num_placeholders: An integer denoting the minimum number of
275
+ placeholders required in the response.
276
+
277
+ Returns:
278
+ A string representing the instruction description.
279
+ """
280
+ self._num_placeholders = num_placeholders
281
+ if self._num_placeholders is None or self._num_placeholders < 0:
282
+ self._num_placeholders = random.randint(1, _NUM_PLACEHOLDERS)
283
+ self._description_pattern = (
284
+ "The response must contain at least {num_placeholders} placeholders "
285
+ + "represented by square brackets, such as [address]."
286
+ )
287
+ return self._description_pattern.format(num_placeholders=self._num_placeholders)
288
+
289
+ def get_instruction_args(self):
290
+ """Returns the keyword args of `build_description`."""
291
+ return {"num_placeholders": self._num_placeholders}
292
+
293
+ def get_instruction_args_keys(self):
294
+ """Returns the args keys of `build_description`."""
295
+ return ["num_placeholders"]
296
+
297
+ def check_following(self, value):
298
+ """Check if the number of placeholders follows the instruction.
299
+
300
+ Args:
301
+ value: A string representing the response.
302
+
303
+ Returns:
304
+ True if the actual number of placeholders in the response is greater than
305
+ or equal to `num_placeholders`; otherwise, False.
306
+ """
307
+ placeholders = re.findall(r"\[.*?\]", value)
308
+ num_placeholders = len(placeholders)
309
+ return num_placeholders >= self._num_placeholders
310
+
311
+
312
+ class BulletListChecker(Instruction):
313
+ """Checks the bullet list in the prompt."""
314
+
315
+ def build_description(self, *, num_bullets=None):
316
+ """Build the instruction description.
317
+
318
+ Args:
319
+ num_bullets: An integer specifying the exact number of bullet lists
320
+ that is required to appear in the response.
321
+
322
+ Returns:
323
+ A string representing the instruction description.
324
+ """
325
+ self._num_bullets = num_bullets
326
+ if self._num_bullets is None or self._num_bullets < 0:
327
+ self._num_bullets = random.randint(1, _NUM_BULLETS)
328
+ self._description_pattern = (
329
+ "Your answer must contain exactly {num_bullets} bullet points. "
330
+ + "Use the markdown bullet points such as:\n"
331
+ + "* This is point 1. \n"
332
+ + "* This is point 2"
333
+ )
334
+ return self._description_pattern.format(num_bullets=self._num_bullets)
335
+
336
+ def get_instruction_args(self):
337
+ """Returns the keyword args of `build_description`."""
338
+ return {"num_bullets": self._num_bullets}
339
+
340
+ def get_instruction_args_keys(self):
341
+ """Returns the args keys of `build_description`."""
342
+ return ["num_bullets"]
343
+
344
+ def check_following(self, value):
345
+ r"""Check if the number of bullet lists meets the requirement.
346
+
347
+ Args:
348
+ value: A string representing the response. The response is expected to
349
+ contain some bullet lists that start with `\*`.
350
+
351
+ Returns:
352
+ True if the actual number of bullet lists in the response meets the
353
+ requirement.
354
+ """
355
+ bullet_lists = re.findall(r"^\s*\*[^\*].*$", value, flags=re.MULTILINE)
356
+ bullet_lists_2 = re.findall(r"^\s*-.*$", value, flags=re.MULTILINE)
357
+ num_bullet_lists = len(bullet_lists) + len(bullet_lists_2)
358
+ return num_bullet_lists == self._num_bullets
359
+
360
+
361
+ class ConstrainedResponseChecker(Instruction):
362
+ """Checks the constrained response."""
363
+
364
+ def build_description(self):
365
+ """Build the instruction description."""
366
+ # A sequence of string(s) representing the options of the expected response.
367
+ self._constrained_responses = _CONSTRAINED_RESPONSE_OPTIONS
368
+ self._description_pattern = (
369
+ "Answer with one of the following options: {response_options}"
370
+ )
371
+ return self._description_pattern.format(
372
+ response_options=self._constrained_responses
373
+ )
374
+
375
+ def get_instruction_args(self):
376
+ """Returns the keyword args of `build_description`."""
377
+ return
378
+
379
+ def get_instruction_args_keys(self):
380
+ """Returns the args keys of `build_description`."""
381
+ return []
382
+
383
+ def check_following(self, value):
384
+ """Checks if the response matches the constrained options.
385
+
386
+ Args:
387
+ value: A string representing the response.
388
+
389
+ Returns:
390
+ True if the actual response contains one of the options in the constrained
391
+ responses; otherwise False.
392
+ """
393
+ value = value.strip()
394
+ for constrained_response in self._constrained_responses:
395
+ if constrained_response in value:
396
+ return True
397
+ return False
398
+
399
+
400
+ # Unused: disabled in instructions_registry.INSTRUCTION_DICT (no dataset rows reference this id).
401
+ class ConstrainedStartChecker(Instruction):
402
+ """Checks the response start."""
403
+
404
+ def build_description(self, *, starter=None):
405
+ """Build the instruction description.
406
+
407
+ Args:
408
+ starter: A string representing the keyword that the response should start
409
+ with.
410
+
411
+ Returns:
412
+ A string representing the instruction description.
413
+ """
414
+ self._starter = starter.strip() if isinstance(starter, str) else starter
415
+ if self._starter is None:
416
+ self._starter = random.choice(_STARTER_OPTIONS)
417
+ self._description_pattern = (
418
+ "During the conversation, when it is your turn, "
419
+ + "please always start with {starter}"
420
+ )
421
+ return self._description_pattern.format(starter=self._starter)
422
+
423
+ def get_instruction_args(self):
424
+ """Returns the keyword args of `build_description`."""
425
+ return {"starter": self._starter}
426
+
427
+ def get_instruction_args_keys(self):
428
+ """Returns the args keys of `build_description`."""
429
+ return ["starter"]
430
+
431
+ def check_following(self, value):
432
+ """Checks if the response starts with the constrained keyword or phrase.
433
+
434
+ Args:
435
+ value: A string representing the response.
436
+
437
+ Returns:
438
+ True if the response starts with the given phrase or keyword that is
439
+ contained in `instruction_args`; otherwise, False.
440
+ """
441
+ response_pattern = r"^\s*" + self._starter + r".*$"
442
+ response_with_constrained_start = re.search(
443
+ response_pattern, value, flags=re.MULTILINE
444
+ )
445
+ return bool(response_with_constrained_start)
446
+
447
+
448
+ class HighlightSectionChecker(Instruction):
449
+ """Checks the highlighted section."""
450
+
451
+ def build_description(self, *, num_highlights=None):
452
+ """Build the instruction description.
453
+
454
+ Args:
455
+ num_highlights: An integer specifying the minimum number of highlighted
456
+ sections.
457
+
458
+ Returns:
459
+ A string representing the instruction description.
460
+ """
461
+ self._num_highlights = num_highlights
462
+ if self._num_highlights is None or self._num_highlights < 0:
463
+ self._num_highlights = random.randint(1, _NUM_HIGHLIGHTED_SECTIONS)
464
+
465
+ self._description_pattern = (
466
+ "Highlight at least {num_highlights} sections in your answer with "
467
+ + "markdown, i.e. *highlighted section*."
468
+ )
469
+
470
+ return self._description_pattern.format(num_highlights=self._num_highlights)
471
+
472
+ def get_instruction_args(self):
473
+ """Returns the keyword args of `build_description`."""
474
+ return {"num_highlights": self._num_highlights}
475
+
476
+ def get_instruction_args_keys(self):
477
+ """Returns the args keys of `build_description`."""
478
+ return ["num_highlights"]
479
+
480
+ def check_following(self, value):
481
+ """Checks if the number of highlighted sections meets the requirement.
482
+
483
+ Args:
484
+ value: a string representing the response. The response is expected to
485
+ contain highlighted sections in the format of *highlighted*.
486
+
487
+ Returns:
488
+ True if the actual number of highlighted sections in the format of
489
+ *highlighted sections* meets the minimum requirement; otherwise False.
490
+ """
491
+ num_highlights = 0
492
+ highlights = re.findall(r"\*[^\n\*]*\*", value)
493
+ double_highlights = re.findall(r"\*\*[^\n\*]*\*\*", value)
494
+ for highlight in highlights:
495
+ if highlight.strip("*").strip():
496
+ num_highlights += 1
497
+ for highlight in double_highlights:
498
+ if highlight.removeprefix("**").removesuffix("**").strip():
499
+ num_highlights += 1
500
+
501
+ return num_highlights >= self._num_highlights
502
+
503
+
504
+ class SectionChecker(Instruction):
505
+ """Checks the sections."""
506
+
507
+ def build_description(self, *, section_spliter=None, num_sections=None):
508
+ """Build the instruction description.
509
+
510
+ Args:
511
+ section_spliter: A string represents the section spliter keyword that
512
+ marks a new section, i.e., `Section` or `SECTION`.
513
+ num_sections: An integer specifying the number of sections.
514
+
515
+ Returns:
516
+ A string representing the instruction description.
517
+ """
518
+ self._section_spliter = (
519
+ section_spliter.strip()
520
+ if isinstance(section_spliter, str)
521
+ else section_spliter
522
+ )
523
+ if self._section_spliter is None:
524
+ self._section_spliter = random.choice(_SECTION_SPLITER)
525
+
526
+ self._num_sections = num_sections
527
+ if self._num_sections is None or self._num_sections < 0:
528
+ self._num_sections = random.randint(1, _NUM_SECTIONS)
529
+
530
+ self._description_pattern = (
531
+ "Your response must have {num_sections} sections. Mark the beginning "
532
+ + "of each section with {section_spliter} X, such as:\n"
533
+ + "{section_spliter} 1\n"
534
+ + "[content of section 1]\n"
535
+ + "{section_spliter} 2\n"
536
+ + "[content of section 2]"
537
+ )
538
+
539
+ return self._description_pattern.format(
540
+ num_sections=self._num_sections, section_spliter=self._section_spliter
541
+ )
542
+
543
+ def get_instruction_args(self):
544
+ """Returns the keyword args of `build_description`."""
545
+ return {
546
+ "section_spliter": self._section_spliter,
547
+ "num_sections": self._num_sections,
548
+ }
549
+
550
+ def get_instruction_args_keys(self):
551
+ """Returns the args keys of `build_description`."""
552
+ return ["section_spliter", "num_sections"]
553
+
554
+ def check_following(self, value):
555
+ """Checks the response contains multiple sections.
556
+
557
+ Args:
558
+ value: A string representing the response. The response is expected
559
+ to contain multiple sections (number of sections is greater than 1).
560
+ A new section starts with `Section 1`, where the number denotes the
561
+ section index.
562
+
563
+ Returns:
564
+ True if the number of sections in the response is greater than or equal to
565
+ the minimum number of sections; otherwise, False.
566
+ """
567
+ section_splitter_patten = r"\s?" + self._section_spliter + r"\s?\d+\s?"
568
+ sections = re.split(section_splitter_patten, value)
569
+ num_sections = len(sections) - 1
570
+ return num_sections >= self._num_sections
571
+
572
+
573
+ class ParagraphChecker(Instruction):
574
+ """Checks the paragraphs."""
575
+
576
+ def build_description(self, *, num_paragraphs=None):
577
+ """Build the instruction description.
578
+
579
+ Args:
580
+ num_paragraphs: An integer specifying the number of paragraphs.
581
+
582
+ Returns:
583
+ A string representing the instruction description.
584
+ """
585
+ self._num_paragraphs = num_paragraphs
586
+ if self._num_paragraphs is None or self._num_paragraphs < 0:
587
+ self._num_paragraphs = random.randint(1, _NUM_PARAGRAPHS)
588
+
589
+ self._description_pattern = (
590
+ "There should be {num_paragraphs} paragraphs. "
591
+ + "Paragraphs are separated with the markdown divider: ***"
592
+ )
593
+
594
+ return self._description_pattern.format(num_paragraphs=self._num_paragraphs)
595
+
596
+ def get_instruction_args(self):
597
+ """Returns the keyword args of `build_description`."""
598
+ return {"num_paragraphs": self._num_paragraphs}
599
+
600
+ def get_instruction_args_keys(self):
601
+ """Returns the args keys of `build_description`."""
602
+ return ["num_paragraphs"]
603
+
604
+ def check_following(self, value):
605
+ """Checks the response contains required number of paragraphs.
606
+
607
+ Args:
608
+ value: A string representing the response. The response may contain
609
+ paragraphs that are separated by the markdown divider: `***`.
610
+
611
+ Returns:
612
+ True if the actual number of paragraphs is the same as required;
613
+ otherwise, False.
614
+ """
615
+ paragraphs = re.split(r"\s?\*\*\*\s?", value)
616
+ num_paragraphs = len(paragraphs)
617
+
618
+ for index, paragraph in enumerate(paragraphs):
619
+ if not paragraph.strip():
620
+ if index == 0 or index == len(paragraphs) - 1:
621
+ num_paragraphs -= 1
622
+ else:
623
+ return False
624
+
625
+ return num_paragraphs == self._num_paragraphs
626
+
627
+
628
+ class PostscriptChecker(Instruction):
629
+ """Checks the postscript."""
630
+
631
+ def build_description(self, *, postscript_marker=None):
632
+ """Build the instruction description.
633
+
634
+ Args:
635
+ postscript_marker: A string containing the keyword that marks the start
636
+ of the postscript section.
637
+
638
+ Returns:
639
+ A string representing the instruction description.
640
+ """
641
+ self._postscript_marker = (
642
+ postscript_marker.strip()
643
+ if isinstance(postscript_marker, str)
644
+ else postscript_marker
645
+ )
646
+ if self._postscript_marker is None:
647
+ self._postscript_marker = random.choice(_POSTSCRIPT_MARKER)
648
+
649
+ self._description_pattern = (
650
+ "At the end of your response, please explicitly add a postscript "
651
+ + "starting with {postscript}"
652
+ )
653
+
654
+ return self._description_pattern.format(postscript=self._postscript_marker)
655
+
656
+ def get_instruction_args(self):
657
+ """Returns the keyword args of `build_description`."""
658
+ return {"postscript_marker": self._postscript_marker}
659
+
660
+ def get_instruction_args_keys(self):
661
+ """Returns the args keys of `build_description`."""
662
+ return ["postscript_marker"]
663
+
664
+ def check_following(self, value):
665
+ """Checks if the response follows the postscript format.
666
+
667
+ Args:
668
+ value: a string representing the response. The response is expected to
669
+ contain a postscript section.
670
+
671
+ Returns:
672
+ True if the response contains a postscript section starting with
673
+ the keyword containing in the `instruction_args`; otherwise False.
674
+ """
675
+ value = value.lower()
676
+ if self._postscript_marker == "P.P.S":
677
+ postscript_pattern = r"\s*p\.\s?p\.\s?s.*$"
678
+ elif self._postscript_marker == "P.S.":
679
+ postscript_pattern = r"\s*p\.\s?s\..*$"
680
+ else:
681
+ postscript_pattern = r"\s*" + self._postscript_marker.lower() + r".*$"
682
+ postscript = re.findall(postscript_pattern, value, flags=re.MULTILINE)
683
+ return bool(postscript)
684
+
685
+
686
+ # Unused: disabled in instructions_registry.INSTRUCTION_DICT (requires an `original_message`
687
+ # kwarg that the published google/IFEval dataset does not provide).
688
+ class RephraseChecker(Instruction):
689
+ """Checks the rephrase."""
690
+
691
+ def build_description(self, *, original_message):
692
+ """Build the instruction description.
693
+
694
+ Args:
695
+ original_message: A string representing the original message. The
696
+ rephrased response should only change its words/sentences in between
697
+ its two asterisks, for example, *change me*. Both original and rephrased
698
+ messages should contain the changes in the form of *change me*.
699
+
700
+ Returns:
701
+ A string representing the instruction description.
702
+ """
703
+ if not self.is_change(original_message):
704
+ raise ValueError(
705
+ f"Message {original_message} does not contain changes "
706
+ "in the form of *change me*."
707
+ )
708
+
709
+ self._reference_without_change = original_message
710
+ self._description = (
711
+ "Rephrasing: Your rephrased response should only"
712
+ + "change the words/sentences in between two asterisks"
713
+ + "such as *change me*."
714
+ )
715
+ return self._description
716
+
717
+ def get_instruction_args(self):
718
+ """Returns the keyword args of `build_description`."""
719
+ return {"original_message": self._reference_without_change}
720
+
721
+ def get_instruction_args_keys(self):
722
+ """Returns the args keys of `build_description`."""
723
+ return ["original_message"]
724
+
725
+ def check_following(self, value):
726
+ r"""Checks if the rephrasing follows the instruction.
727
+
728
+ Args:
729
+ value: A string representing the response, which is expected to rephras
730
+ the string of `instruction_args`.
731
+
732
+ Returns:
733
+ True if `value` and `instruction_args` only differ by the words/sentences
734
+ in between two asterisks such as *change me*; otherwise, False.
735
+ """
736
+ if not self.is_change(value):
737
+ raise ValueError(
738
+ f"value {value} does not contain changes in the form of *change me*."
739
+ )
740
+
741
+ response_without_changes = self.strip_changes(value)
742
+ reference_without_changes = self.strip_changes(self._reference_without_change)
743
+
744
+ return response_without_changes == reference_without_changes
745
+
746
+ def is_change(self, response):
747
+ """Check if there is change in the response in the form of *change me*."""
748
+ return re.search(r"\*.*\*", response)
749
+
750
+ def strip_changes(self, response):
751
+ """Strips off the changes."""
752
+ return re.sub(r"\*.*?\*", "", response)
753
+
754
+
755
+ class KeywordChecker(Instruction):
756
+ """Check the existence of certain keywords."""
757
+
758
+ def build_description(self, *, keywords=None):
759
+ """Build the instruction description.
760
+
761
+ Args:
762
+ keywords: A sequence of strings representing the keywords that are
763
+ expected in the response.
764
+
765
+ Returns:
766
+ A string representing the instruction description.
767
+ """
768
+ if not keywords:
769
+ self._keywords = instructions_util.generate_keywords(
770
+ num_keywords=_NUM_KEYWORDS
771
+ )
772
+ else:
773
+ self._keywords = keywords
774
+ self._keywords = sorted(self._keywords)
775
+
776
+ self._description_pattern = "Include keywords {keywords} in the response."
777
+
778
+ return self._description_pattern.format(keywords=self._keywords)
779
+
780
+ def get_instruction_args(self):
781
+ """Returns the keyword args of `build_description`."""
782
+ return {"keywords": self._keywords}
783
+
784
+ def get_instruction_args_keys(self):
785
+ """Returns the args keys of `build_description`."""
786
+ return ["keywords"]
787
+
788
+ def check_following(self, value):
789
+ """Check if the response contain the expected keywords."""
790
+ for keyword in self._keywords:
791
+ if not re.search(keyword, value, flags=re.IGNORECASE):
792
+ return False
793
+ return True
794
+
795
+
796
+ class KeywordFrequencyChecker(Instruction):
797
+ """Check the keyword frequency."""
798
+
799
+ def build_description(self, *, keyword=None, frequency=None, relation=None):
800
+ """Build the instruction description.
801
+
802
+ Args:
803
+ keyword: A string representing a keyword that is expected in the response.
804
+ frequency: An integer specifying the number of times `keyword` is expected
805
+ to appear in the response.
806
+ relation: A string in (`less than`, `at least`), defining the relational
807
+ operator for comparison.
808
+ Two relational comparisons are supported for now:
809
+ if 'less than', the actual number of occurrences < frequency;
810
+ if 'at least', the actual number of occurrences >= frequency.
811
+
812
+ Returns:
813
+ A string representing the instruction description.
814
+ """
815
+ if not keyword:
816
+ self._keyword = instructions_util.generate_keywords(num_keywords=1)[0]
817
+ else:
818
+ self._keyword = keyword.strip()
819
+
820
+ self._frequency = frequency
821
+ if self._frequency is None or self._frequency < 0:
822
+ self._frequency = random.randint(1, _KEYWORD_FREQUENCY)
823
+
824
+ if relation is None:
825
+ self._comparison_relation = random.choice(_COMPARISON_RELATION)
826
+ elif relation not in _COMPARISON_RELATION:
827
+ raise ValueError(
828
+ "The supported relation for comparison must be in "
829
+ f"{_COMPARISON_RELATION}, but {relation} is given."
830
+ )
831
+ else:
832
+ self._comparison_relation = relation
833
+
834
+ self._description_pattern = (
835
+ "In your response, the word {keyword} should appear {relation} "
836
+ + "{frequency} times."
837
+ )
838
+
839
+ return self._description_pattern.format(
840
+ keyword=self._keyword,
841
+ relation=self._comparison_relation,
842
+ frequency=self._frequency,
843
+ )
844
+
845
+ def get_instruction_args(self):
846
+ """Returns the keyword args of `build_description`."""
847
+ return {
848
+ "keyword": self._keyword,
849
+ "frequency": self._frequency,
850
+ "relation": self._comparison_relation,
851
+ }
852
+
853
+ def get_instruction_args_keys(self):
854
+ """Returns the args keys of `build_description`."""
855
+ return ["keyword", "frequency", "relation"]
856
+
857
+ def check_following(self, value):
858
+ """Checks if the response contain the keyword with required frequency."""
859
+ actual_occurrences = len(re.findall(self._keyword, value, flags=re.IGNORECASE))
860
+
861
+ if self._comparison_relation == _COMPARISON_RELATION[0]:
862
+ return actual_occurrences < self._frequency
863
+ elif self._comparison_relation == _COMPARISON_RELATION[1]:
864
+ return actual_occurrences >= self._frequency
865
+
866
+
867
+ class NumberOfWords(Instruction):
868
+ """Checks the number of words."""
869
+
870
+ def build_description(self, *, num_words=None, relation=None):
871
+ """Build the instruction description.
872
+
873
+ Args:
874
+ num_words: An integer specifying the number of words contained in the
875
+ response.
876
+ relation: A string in (`less than`, `at least`), defining the relational
877
+ operator for comparison.
878
+ Two relational comparisons are supported for now:
879
+ if 'less than', the actual number of words < num_words;
880
+ if 'at least', the actual number of words >= num_words.
881
+
882
+ Returns:
883
+ A string representing the instruction description.
884
+ """
885
+ self._num_words = num_words
886
+ if self._num_words is None or self._num_words < 0:
887
+ self._num_words = random.randint(
888
+ _NUM_WORDS_LOWER_LIMIT, _NUM_WORDS_UPPER_LIMIT
889
+ )
890
+
891
+ if relation is None:
892
+ self._comparison_relation = random.choice(_COMPARISON_RELATION)
893
+ elif relation not in _COMPARISON_RELATION:
894
+ raise ValueError(
895
+ "The supported relation for comparison must be in "
896
+ f"{_COMPARISON_RELATION}, but {relation} is given."
897
+ )
898
+ else:
899
+ self._comparison_relation = relation
900
+
901
+ self._description_pattern = "Answer with {relation} {num_words} words."
902
+
903
+ return self._description_pattern.format(
904
+ relation=self._comparison_relation, num_words=self._num_words
905
+ )
906
+
907
+ def get_instruction_args(self):
908
+ """Returns the keyword args of `build_description`."""
909
+ return {"num_words": self._num_words, "relation": self._comparison_relation}
910
+
911
+ def get_instruction_args_keys(self):
912
+ """Returns the args keys of `build_description`."""
913
+ return ["num_words", "relation"]
914
+
915
+ def check_following(self, value):
916
+ """Checks if the response contains the expected number of words."""
917
+ num_words = instructions_util.count_words(value)
918
+
919
+ if self._comparison_relation == _COMPARISON_RELATION[0]:
920
+ return num_words < self._num_words
921
+ elif self._comparison_relation == _COMPARISON_RELATION[1]:
922
+ return num_words >= self._num_words
923
+
924
+
925
+ class JsonFormat(Instruction):
926
+ """Check the Json format."""
927
+
928
+ def build_description(self):
929
+ self._description_pattern = (
930
+ "Entire output should be wrapped in JSON format. You can use markdown"
931
+ " ticks such as ```."
932
+ )
933
+ return self._description_pattern
934
+
935
+ def get_instruction_args(self):
936
+ """Returns the keyword args of `build_description`."""
937
+ return
938
+
939
+ def get_instruction_args_keys(self):
940
+ """Returns the args keys of `build_description`."""
941
+ return []
942
+
943
+ def check_following(self, value):
944
+ value = (
945
+ value.strip()
946
+ .removeprefix("```json")
947
+ .removeprefix("```Json")
948
+ .removeprefix("```JSON")
949
+ .removeprefix("```")
950
+ .removesuffix("```")
951
+ .strip()
952
+ )
953
+ try:
954
+ json.loads(value)
955
+ except ValueError:
956
+ return False
957
+ return True
958
+
959
+
960
+ class ParagraphFirstWordCheck(Instruction):
961
+ """Check the paragraph and the first word of the nth paragraph."""
962
+
963
+ def build_description(
964
+ self, num_paragraphs=None, nth_paragraph=None, first_word=None
965
+ ):
966
+ r"""Build the instruction description.
967
+
968
+ Args:
969
+ num_paragraphs: An integer indicating the number of paragraphs expected
970
+ in the response. A paragraph is a subset of the string that is
971
+ expected to be separated by '\n\n'.
972
+ nth_paragraph: An integer indicating the paragraph number that we look at.
973
+ Note that n starts from 1.
974
+ first_word: A string that represent the first word of the bth paragraph.
975
+
976
+ Returns:
977
+ A string representing the instruction description.
978
+ """
979
+ self._num_paragraphs = num_paragraphs
980
+ if self._num_paragraphs is None or self._num_paragraphs < 0:
981
+ self._num_paragraphs = random.randint(1, _NUM_PARAGRAPHS)
982
+
983
+ self._nth_paragraph = nth_paragraph
984
+ if (
985
+ self._nth_paragraph is None
986
+ or self._nth_paragraph <= 0
987
+ or self._nth_paragraph > self._num_paragraphs
988
+ ):
989
+ self._nth_paragraph = random.randint(1, self._num_paragraphs + 1)
990
+
991
+ self._first_word = first_word
992
+ if self._first_word is None:
993
+ self._first_word = instructions_util.generate_keywords(num_keywords=1)[0]
994
+ self._first_word = self._first_word.lower()
995
+
996
+ self._description_pattern = (
997
+ "There should be {num_paragraphs} paragraphs. "
998
+ + "Paragraphs and only paragraphs are separated with each other by two "
999
+ + "new lines as if it was '\\n\\n' in python. "
1000
+ + "Paragraph {nth_paragraph} must start with word {first_word}."
1001
+ )
1002
+
1003
+ return self._description_pattern.format(
1004
+ num_paragraphs=self._num_paragraphs,
1005
+ nth_paragraph=self._nth_paragraph,
1006
+ first_word=self._first_word,
1007
+ )
1008
+
1009
+ def get_instruction_args(self):
1010
+ """Returns the keyword args of `build_description`."""
1011
+ return {
1012
+ "num_paragraphs": self._num_paragraphs,
1013
+ "nth_paragraph": self._nth_paragraph,
1014
+ "first_word": self._first_word,
1015
+ }
1016
+
1017
+ def get_instruction_args_keys(self):
1018
+ """Returns the args keys of `build_description`."""
1019
+ return ["num_paragraphs", "nth_paragraph", "first_word"]
1020
+
1021
+ def check_following(self, value):
1022
+ """Checks for required number of paragraphs and correct first word.
1023
+
1024
+ Args:
1025
+ value: a string representing the response. The response may contain
1026
+ paragraphs that are separated by two new lines and the first word of
1027
+ the nth paragraph will have to match a specified word.
1028
+
1029
+ Returns:
1030
+ True if the number of paragraphs is the same as required and the first
1031
+ word of the specified paragraph is the same as required. Otherwise, false.
1032
+ """
1033
+ paragraphs = re.split(r"\n\n", value)
1034
+ num_paragraphs = len(paragraphs)
1035
+
1036
+ for paragraph in paragraphs:
1037
+ if not paragraph.strip():
1038
+ num_paragraphs -= 1
1039
+
1040
+ # check that index doesn't go out of bounds
1041
+ if self._nth_paragraph <= num_paragraphs:
1042
+ paragraph = paragraphs[self._nth_paragraph - 1].strip()
1043
+ if not paragraph:
1044
+ return False
1045
+ else:
1046
+ return False
1047
+
1048
+ first_word = ""
1049
+ punctuation = {".", ",", "?", "!", "'", '"'}
1050
+
1051
+ # get first word and remove punctuation
1052
+ word = paragraph.split()[0].strip()
1053
+ # TODO(jeffrey): make more complex?
1054
+ word = word.lstrip("'")
1055
+ word = word.lstrip('"')
1056
+
1057
+ for letter in word:
1058
+ if letter in punctuation:
1059
+ break
1060
+ first_word += letter.lower()
1061
+
1062
+ return num_paragraphs == self._num_paragraphs and first_word == self._first_word
1063
+
1064
+
1065
+ # TODO(jeffrey) add relation - at least/at most?
1066
+ # Unused: disabled in instructions_registry.INSTRUCTION_DICT (no dataset rows reference this id).
1067
+ class KeySentenceChecker(Instruction):
1068
+ """Check the existence of certain key sentences."""
1069
+
1070
+ def build_description(self, key_sentences=None, num_sentences=None):
1071
+ """Build the instruction description.
1072
+
1073
+ Args:
1074
+ key_sentences: A sequences of strings representing the key sentences that
1075
+ are expected in the response.
1076
+ num_sentences: The number of key sentences that are expected to be seen in
1077
+ the response.
1078
+
1079
+ Returns:
1080
+ A string representing the instruction description.
1081
+ """
1082
+ if not key_sentences:
1083
+ # TODO(jeffrey) make a generate sentences function? wonderwords package
1084
+ self._key_sentences = {"For now, this is fine."}
1085
+ else:
1086
+ self._key_sentences = key_sentences
1087
+
1088
+ if not num_sentences:
1089
+ self._num_sentences = random.randint(1, len(self._key_sentences))
1090
+ else:
1091
+ self._num_sentences = num_sentences
1092
+
1093
+ self._description_pattern = (
1094
+ "Include {num_sentences} of the following sentences {key_sentences}"
1095
+ )
1096
+
1097
+ return self._description_pattern.format(
1098
+ num_sentences=self._num_sentences, key_sentences=self._key_sentences
1099
+ )
1100
+
1101
+ def get_instruction_args(self):
1102
+ """Returns the keyword args of `build_description`."""
1103
+ return {
1104
+ "num_sentences": self._num_sentences,
1105
+ "key_sentences": list(self._key_sentences),
1106
+ }
1107
+
1108
+ def get_instruction_args_keys(self):
1109
+ """Returns the args keys of `build_description`."""
1110
+ return ["num_sentences", "key_sentences"]
1111
+
1112
+ def check_following(self, value):
1113
+ """Checks if the response contains the expected key sentences."""
1114
+ count = 0
1115
+ sentences = instructions_util.split_into_sentences(value)
1116
+ for sentence in self._key_sentences:
1117
+ if sentence in sentences:
1118
+ count += 1
1119
+
1120
+ return count == self._num_sentences
1121
+
1122
+
1123
+ class ForbiddenWords(Instruction):
1124
+ """Checks that specified words are not used in response."""
1125
+
1126
+ def build_description(self, forbidden_words=None):
1127
+ """Build the instruction description.
1128
+
1129
+ Args:
1130
+ forbidden_words: A sequences of strings representing words that are not
1131
+ allowed in the response.
1132
+
1133
+ Returns:
1134
+ A string representing the instruction description.
1135
+ """
1136
+ if not forbidden_words:
1137
+ self._forbidden_words = instructions_util.generate_keywords(
1138
+ num_keywords=_NUM_KEYWORDS
1139
+ )
1140
+ else:
1141
+ self._forbidden_words = list(set(forbidden_words))
1142
+ self._forbidden_words = sorted(self._forbidden_words)
1143
+ self._description_pattern = (
1144
+ "Do not include keywords {forbidden_words} in the response."
1145
+ )
1146
+
1147
+ return self._description_pattern.format(forbidden_words=self._forbidden_words)
1148
+
1149
+ def get_instruction_args(self):
1150
+ """Returns the keyword args of `build_description`."""
1151
+ return {"forbidden_words": self._forbidden_words}
1152
+
1153
+ def get_instruction_args_keys(self):
1154
+ """Returns the args keys of `build_description`."""
1155
+ return ["forbidden_words"]
1156
+
1157
+ def check_following(self, value):
1158
+ """Check if the response does not contain the expected keywords."""
1159
+ for word in self._forbidden_words:
1160
+ if re.search(r"\b" + word + r"\b", value, flags=re.IGNORECASE):
1161
+ return False
1162
+ return True
1163
+
1164
+
1165
+ # Unused: disabled in instructions_registry.INSTRUCTION_DICT (no dataset rows reference this id).
1166
+ class RephraseParagraph(Instruction):
1167
+ """Checks that the paragraph is rephrased."""
1168
+
1169
+ def build_description(self, *, original_paragraph, low, high):
1170
+ """Builds the instruction description.
1171
+
1172
+ Args:
1173
+ original_paragraph: A string presenting the original paragraph. The
1174
+ rephrases response should have between low-high words in common.
1175
+ low: An integer presenting the lower bound of similar words.
1176
+ high: An integer representing the upper bound of similar words.
1177
+
1178
+ Returns:
1179
+ A string representing the instruction description.
1180
+ """
1181
+ # TODO(jeffrey) make more encompassing
1182
+ self._original_paragraph = original_paragraph
1183
+ self._low = low
1184
+ self._high = high
1185
+
1186
+ self._description = (
1187
+ "Rephrase the following paragraph: "
1188
+ + "{original_paragraph}\nYour response should have "
1189
+ + "between {low} and {high} of the same words. "
1190
+ + "Words are the same if and only if all of the "
1191
+ + "letters, ignoring cases, are the same. For "
1192
+ + "example, 'run' is the same as 'Run' but different "
1193
+ + "to 'ran'."
1194
+ )
1195
+
1196
+ return self._description.format(
1197
+ original_paragraph=original_paragraph, low=self._low, high=self._high
1198
+ )
1199
+
1200
+ def get_instruction_args(self):
1201
+ """Returns the keyword args of `build_description`."""
1202
+ return {
1203
+ "original_paragraph": self._original_paragraph,
1204
+ "low": self._low,
1205
+ "high": self._high,
1206
+ }
1207
+
1208
+ def get_instruction_args_keys(self):
1209
+ """Returns the args keys of `build_description`."""
1210
+ return ["original_paragraph", "low", "high"]
1211
+
1212
+ def check_following(self, value):
1213
+ val_words = re.findall(r"\w+", value.lower())
1214
+ original_words = re.findall(r"\w+", self._original_paragraph.lower())
1215
+ similar_words = 0
1216
+
1217
+ dict_val = collections.Counter(val_words)
1218
+ dict_original = collections.Counter(original_words)
1219
+
1220
+ for word in dict_original:
1221
+ similar_words += min(dict_original[word], dict_val[word])
1222
+
1223
+ return similar_words >= self._low and similar_words <= self._high
1224
+
1225
+
1226
+ class TwoResponsesChecker(Instruction):
1227
+ """Check that two responses were given."""
1228
+
1229
+ def build_description(self):
1230
+ """Build the instruction description."""
1231
+ self._description_pattern = (
1232
+ "Give two different responses. Responses and only responses should"
1233
+ " be separated by 6 asterisk symbols: ******."
1234
+ )
1235
+ return self._description_pattern
1236
+
1237
+ def get_instruction_args(self):
1238
+ """Returns the keyword args of `build_description`."""
1239
+ return
1240
+
1241
+ def get_instruction_args_keys(self):
1242
+ """Returns the args keys of `build_description`."""
1243
+ return []
1244
+
1245
+ def check_following(self, value):
1246
+ """Checks if the response has two different answers.
1247
+
1248
+ Args:
1249
+ value: A string representing the response.
1250
+
1251
+ Returns:
1252
+ True if two responses are detected and false otherwise.
1253
+ """
1254
+ valid_responses = []
1255
+ responses = value.split("******")
1256
+ for index, response in enumerate(responses):
1257
+ if not response.strip():
1258
+ if index != 0 and index != len(responses) - 1:
1259
+ return False
1260
+ else:
1261
+ valid_responses.append(response)
1262
+ return (
1263
+ len(valid_responses) == 2
1264
+ and valid_responses[0].strip() != valid_responses[1].strip()
1265
+ )
1266
+
1267
+
1268
+ class RepeatPromptThenAnswer(Instruction):
1269
+ """Checks that Prompt is first repeated then answered."""
1270
+
1271
+ def build_description(self, *, prompt_to_repeat=None):
1272
+ """Build the instruction description.
1273
+
1274
+ Args:
1275
+ prompt_to_repeat: The prompt that is meant to be repeated.
1276
+
1277
+ Returns:
1278
+ A string representing the instruction description.
1279
+ """
1280
+ if not prompt_to_repeat:
1281
+ raise ValueError("prompt_to_repeat must be set.")
1282
+ else:
1283
+ self._prompt_to_repeat = prompt_to_repeat
1284
+ self._description_pattern = (
1285
+ "First repeat the request word for word without change,"
1286
+ " then give your answer (1. do not say any words or characters"
1287
+ " before repeating the request; 2. the request you need to repeat"
1288
+ " does not include this sentence)"
1289
+ )
1290
+ return self._description_pattern
1291
+
1292
+ def get_instruction_args(self):
1293
+ return {"prompt_to_repeat": self._prompt_to_repeat}
1294
+
1295
+ def get_instruction_args_keys(self):
1296
+ """Returns the args keys of `build_description`."""
1297
+ return ["prompt_to_repeat"]
1298
+
1299
+ def check_following(self, value):
1300
+ return bool(
1301
+ value.strip().lower().startswith(self._prompt_to_repeat.strip().lower())
1302
+ )
1303
+
1304
+
1305
+ class EndChecker(Instruction):
1306
+ """Checks that the prompt ends with a given phrase."""
1307
+
1308
+ def build_description(self, *, end_phrase=None):
1309
+ """Build the instruction description.
1310
+
1311
+ Args:
1312
+ end_phrase: A string representing the phrase the response should end with.
1313
+
1314
+ Returns:
1315
+ A string representing the instruction description.
1316
+ """
1317
+ self._end_phrase = (
1318
+ end_phrase.strip() if isinstance(end_phrase, str) else end_phrase
1319
+ )
1320
+ if self._end_phrase is None:
1321
+ self._end_phrase = random.choice(_ENDING_OPTIONS)
1322
+ self._description_pattern = (
1323
+ "Finish your response with this exact phrase {ender}. "
1324
+ "No other words should follow this phrase."
1325
+ )
1326
+ return self._description_pattern.format(ender=self._end_phrase)
1327
+
1328
+ def get_instruction_args(self):
1329
+ return {"end_phrase": self._end_phrase}
1330
+
1331
+ def get_instruction_args_keys(self):
1332
+ """Returns the args keys of `build_description`."""
1333
+ return ["end_phrase"]
1334
+
1335
+ def check_following(self, value):
1336
+ """Checks if the response ends with the expected phrase."""
1337
+ value = value.strip().strip('"').lower()
1338
+ self._end_phrase = self._end_phrase.strip().lower()
1339
+ return value.endswith(self._end_phrase)
1340
+
1341
+
1342
+ class TitleChecker(Instruction):
1343
+ """Checks the response for a title."""
1344
+
1345
+ def build_description(self):
1346
+ """Build the instruction description."""
1347
+ self._description_pattern = (
1348
+ "Your answer must contain a title, wrapped in double angular brackets,"
1349
+ " such as <<poem of joy>>."
1350
+ )
1351
+ return self._description_pattern
1352
+
1353
+ def get_instruction_args(self):
1354
+ return None
1355
+
1356
+ def get_instruction_args_keys(self):
1357
+ """Returns the args keys of `build_description`."""
1358
+ return []
1359
+
1360
+ def check_following(self, value):
1361
+ """Checks if the response contains a title."""
1362
+ pattern = r"<<[^\n]+>>"
1363
+ re_pattern = re.compile(pattern)
1364
+ titles = re.findall(re_pattern, value)
1365
+
1366
+ return any(title.lstrip("<").rstrip(">").strip() for title in titles)
1367
+
1368
+
1369
+ class LetterFrequencyChecker(Instruction):
1370
+ """Checks letter frequency."""
1371
+
1372
+ def build_description(self, *, letter=None, let_frequency=None, let_relation=None):
1373
+ """Build the instruction description.
1374
+
1375
+ Args:
1376
+ letter: A string representing a letter that is expected in the response.
1377
+ let_frequency: An integer specifying the number of times `keyword` is
1378
+ expected to appear in the response.
1379
+ let_relation: A string in (`less than`, `at least`), defining the
1380
+ relational operator for comparison. Two relational comparisons are
1381
+ supported for now; if 'less than', the actual number of
1382
+ occurrences < frequency; if 'at least', the actual number of
1383
+ occurrences >= frequency.
1384
+
1385
+ Returns:
1386
+ A string representing the instruction description.
1387
+ """
1388
+ if (
1389
+ not letter
1390
+ or len(letter) > 1
1391
+ or ord(letter.lower()) < 97
1392
+ or ord(letter.lower()) > 122
1393
+ ):
1394
+ self._letter = random.choice(list(string.ascii_letters))
1395
+ else:
1396
+ self._letter = letter.strip()
1397
+ self._letter = self._letter.lower()
1398
+
1399
+ self._frequency = let_frequency
1400
+ if self._frequency is None or self._frequency < 0:
1401
+ self._frequency = random.randint(1, _LETTER_FREQUENCY)
1402
+
1403
+ if let_relation is None:
1404
+ self._comparison_relation = random.choice(_COMPARISON_RELATION)
1405
+ elif let_relation not in _COMPARISON_RELATION:
1406
+ raise ValueError(
1407
+ "The supported relation for comparison must be in "
1408
+ f"{_COMPARISON_RELATION}, but {let_relation} is given."
1409
+ )
1410
+ else:
1411
+ self._comparison_relation = let_relation
1412
+
1413
+ self._description_pattern = (
1414
+ "In your response, the letter {letter} should appear {let_relation}"
1415
+ " {let_frequency} times."
1416
+ )
1417
+
1418
+ return self._description_pattern.format(
1419
+ letter=self._letter,
1420
+ let_frequency=self._frequency,
1421
+ let_relation=self._comparison_relation,
1422
+ )
1423
+
1424
+ def get_instruction_args(self):
1425
+ """Returns the keyword args of build description."""
1426
+ return {
1427
+ "letter": self._letter,
1428
+ "let_frequency": self._frequency,
1429
+ "let_relation": self._comparison_relation,
1430
+ }
1431
+
1432
+ def get_instruction_args_keys(self):
1433
+ """Returns the args keys of `build_description`."""
1434
+ return ["letter", "let_frequency", "let_relation"]
1435
+
1436
+ def check_following(self, value):
1437
+ """Checks that the response contains the letter at the right frequency."""
1438
+ value = value.lower()
1439
+ letters = collections.Counter(value)
1440
+
1441
+ if self._comparison_relation == _COMPARISON_RELATION[0]:
1442
+ return letters[self._letter] < self._frequency
1443
+ else:
1444
+ return letters[self._letter] >= self._frequency
1445
+
1446
+
1447
+ class CapitalLettersEnglishChecker(Instruction):
1448
+ """Checks that the response is in english and is in all capital letters."""
1449
+
1450
+ def build_description(self):
1451
+ """Build the instruction description."""
1452
+ self._description_pattern = (
1453
+ "Your entire response should be in English, and in all capital letters."
1454
+ )
1455
+ return self._description_pattern
1456
+
1457
+ def get_instruction_args(self):
1458
+ return None
1459
+
1460
+ def get_instruction_args_keys(self):
1461
+ """Returns the args keys of `build_description`."""
1462
+ return []
1463
+
1464
+ def check_following(self, value):
1465
+ """Checks that the response is in English and in all capital letters."""
1466
+ assert isinstance(value, str)
1467
+
1468
+ try:
1469
+ return value.isupper() and langdetect.detect(value) == "en"
1470
+ except langdetect.LangDetectException as e:
1471
+ # Count as instruction is followed.
1472
+ logger.error(
1473
+ "Unable to detect language for text %s due to %s", value, e
1474
+ ) # refex: disable=pytotw.037
1475
+ return True
1476
+
1477
+
1478
+ class LowercaseLettersEnglishChecker(Instruction):
1479
+ """Checks that the response is in english and is in all lowercase letters."""
1480
+
1481
+ def build_description(self):
1482
+ """Build the instruction description."""
1483
+ self._description_pattern = (
1484
+ "Your entire response should be in English, and in all lowercase"
1485
+ " letters. No capital letters are allowed."
1486
+ )
1487
+ return self._description_pattern
1488
+
1489
+ def get_instruction_args(self):
1490
+ return None
1491
+
1492
+ def get_instruction_args_keys(self):
1493
+ """Returns the args keys of `build_description`."""
1494
+ return []
1495
+
1496
+ def check_following(self, value):
1497
+ """Checks that the response is in English and in all lowercase letters."""
1498
+ assert isinstance(value, str)
1499
+
1500
+ try:
1501
+ return value.islower() and langdetect.detect(value) == "en"
1502
+ except langdetect.LangDetectException as e:
1503
+ # Count as instruction is followed.
1504
+ logger.error(
1505
+ "Unable to detect language for text %s due to %s", value, e
1506
+ ) # refex: disable=pytotw.037
1507
+ return True
1508
+
1509
+
1510
+ class CommaChecker(Instruction):
1511
+ """Checks the response for no commas."""
1512
+
1513
+ def build_description(self):
1514
+ """Build the instruction description."""
1515
+ self._description_pattern = (
1516
+ "In your entire response, refrain from the use of any commas."
1517
+ )
1518
+ return self._description_pattern
1519
+
1520
+ def get_instruction_args(self):
1521
+ return None
1522
+
1523
+ def get_instruction_args_keys(self):
1524
+ """Returns the args keys of `build_description`."""
1525
+ return []
1526
+
1527
+ def check_following(self, value):
1528
+ """Checks that the response does not contain commas."""
1529
+ return not re.search(r"\,", value)
1530
+
1531
+
1532
+ class CapitalWordFrequencyChecker(Instruction):
1533
+ """Checks frequency of words with all capital letters."""
1534
+
1535
+ def build_description(
1536
+ self,
1537
+ capital_frequency=None,
1538
+ capital_relation=None,
1539
+ ):
1540
+ """Build the instruction description.
1541
+
1542
+ Args:
1543
+ capital_frequency: An integer that represents the number of words that
1544
+ should be in all capital letters.
1545
+ capital_relation: A string that is 'at least' or 'at most' that refers to
1546
+ the frequency.
1547
+
1548
+ Returns:
1549
+ A string representing the instruction description.
1550
+ """
1551
+ self._frequency = capital_frequency
1552
+ if self._frequency is None:
1553
+ self._frequency = random.randint(1, _ALL_CAPITAL_WORD_FREQUENCY)
1554
+
1555
+ self._comparison_relation = capital_relation
1556
+ if capital_relation is None:
1557
+ self._comparison_relation = random.choice(_COMPARISON_RELATION)
1558
+ elif capital_relation not in _COMPARISON_RELATION:
1559
+ raise ValueError(
1560
+ "The supported relation for comparison must be in "
1561
+ f"{_COMPARISON_RELATION}, but {capital_relation} is given."
1562
+ )
1563
+
1564
+ self._description_pattern = (
1565
+ "In your response, words with all capital letters should appear"
1566
+ " {relation} {frequency} times."
1567
+ )
1568
+
1569
+ return self._description_pattern.format(
1570
+ frequency=self._frequency, relation=self._comparison_relation
1571
+ )
1572
+
1573
+ def get_instruction_args(self):
1574
+ """Returns the keyword args of build description."""
1575
+ return {
1576
+ "capital_frequency": self._frequency,
1577
+ "capital_relation": self._comparison_relation,
1578
+ }
1579
+
1580
+ def get_instruction_args_keys(self):
1581
+ """Returns the args keys of `build_description`."""
1582
+ return ["capital_frequency", "capital_relation"]
1583
+
1584
+ def check_following(self, value):
1585
+ """Checks the frequency of words with all capital letters."""
1586
+ # Hyphenated words will count as one word
1587
+ words = instructions_util.nltk.word_tokenize(value)
1588
+ capital_words = [word for word in words if word.isupper()]
1589
+
1590
+ capital_words = len(capital_words)
1591
+
1592
+ if self._comparison_relation == _COMPARISON_RELATION[0]:
1593
+ return capital_words < self._frequency
1594
+ else:
1595
+ return capital_words >= self._frequency
1596
+
1597
+
1598
+ class QuotationChecker(Instruction):
1599
+ """Checks response is wrapped with double quotation marks."""
1600
+
1601
+ def build_description(self):
1602
+ """Build the instruction description."""
1603
+ self._description_pattern = (
1604
+ "Wrap your entire response with double quotation marks."
1605
+ )
1606
+ return self._description_pattern
1607
+
1608
+ def get_instruction_args(self):
1609
+ """Returns the keyword args of build description."""
1610
+ return
1611
+
1612
+ def get_instruction_args_keys(self):
1613
+ """Returns the args keys of `build_description`."""
1614
+ return []
1615
+
1616
+ def check_following(self, value):
1617
+ """Checks if the response is wrapped with double quotation marks."""
1618
+ value = value.strip()
1619
+ return len(value) > 1 and value[0] == '"' and value[-1] == '"'
eval/ifeval/instructions_registry.py ADDED
@@ -0,0 +1,168 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 The Google Research Authors.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ """Registry of all instructions."""
16
+
17
+ from . import instructions
18
+
19
+
20
+ _KEYWORD = "keywords:"
21
+
22
+ _LANGUAGE = "language:"
23
+
24
+ _LENGTH = "length_constraints:"
25
+
26
+ _CONTENT = "detectable_content:"
27
+
28
+ _FORMAT = "detectable_format:"
29
+
30
+ _MULTITURN = "multi-turn:"
31
+
32
+ _COMBINATION = "combination:"
33
+
34
+ _STARTEND = "startend:"
35
+
36
+ _CHANGE_CASES = "change_case:"
37
+
38
+ _PUNCTUATION = "punctuation:"
39
+
40
+ INSTRUCTION_DICT = {
41
+ _KEYWORD + "existence": instructions.KeywordChecker,
42
+ _KEYWORD + "frequency": instructions.KeywordFrequencyChecker,
43
+ # TODO(jeffreyzhou): make a proper set of sentences to choose from
44
+ # _KEYWORD + "key_sentences": instructions.KeySentenceChecker,
45
+ _KEYWORD + "forbidden_words": instructions.ForbiddenWords,
46
+ _KEYWORD + "letter_frequency": instructions.LetterFrequencyChecker,
47
+ _LANGUAGE + "response_language": instructions.ResponseLanguageChecker,
48
+ _LENGTH + "number_sentences": instructions.NumberOfSentences,
49
+ _LENGTH + "number_paragraphs": instructions.ParagraphChecker,
50
+ _LENGTH + "number_words": instructions.NumberOfWords,
51
+ _LENGTH + "nth_paragraph_first_word": instructions.ParagraphFirstWordCheck,
52
+ _CONTENT + "number_placeholders": instructions.PlaceholderChecker,
53
+ _CONTENT + "postscript": instructions.PostscriptChecker,
54
+ _FORMAT + "number_bullet_lists": instructions.BulletListChecker,
55
+ # TODO(jeffreyzhou): Pre-create paragraph or use prompt to replace
56
+ # _CONTENT + "rephrase_paragraph": instructions.RephraseParagraph,
57
+ _FORMAT + "constrained_response": instructions.ConstrainedResponseChecker,
58
+ _FORMAT + "number_highlighted_sections": (instructions.HighlightSectionChecker),
59
+ _FORMAT + "multiple_sections": instructions.SectionChecker,
60
+ # TODO(tianjianlu): Re-enable rephrasing with preprocessing the message.
61
+ # _FORMAT + "rephrase": instructions.RephraseChecker,
62
+ _FORMAT + "json_format": instructions.JsonFormat,
63
+ _FORMAT + "title": instructions.TitleChecker,
64
+ # TODO(tianjianlu): Re-enable with specific prompts.
65
+ # _MULTITURN + "constrained_start": instructions.ConstrainedStartChecker,
66
+ _COMBINATION + "two_responses": instructions.TwoResponsesChecker,
67
+ _COMBINATION + "repeat_prompt": instructions.RepeatPromptThenAnswer,
68
+ _STARTEND + "end_checker": instructions.EndChecker,
69
+ _CHANGE_CASES + "capital_word_frequency": instructions.CapitalWordFrequencyChecker,
70
+ _CHANGE_CASES + "english_capital": instructions.CapitalLettersEnglishChecker,
71
+ _CHANGE_CASES + "english_lowercase": instructions.LowercaseLettersEnglishChecker,
72
+ _PUNCTUATION + "no_comma": instructions.CommaChecker,
73
+ _STARTEND + "quotation": instructions.QuotationChecker,
74
+ }
75
+
76
+ INSTRUCTION_CONFLICTS = {
77
+ _KEYWORD + "existence": {_KEYWORD + "existence"},
78
+ _KEYWORD + "frequency": {_KEYWORD + "frequency"},
79
+ # TODO(jeffreyzhou): make a proper set of sentences to choose from
80
+ # _KEYWORD + "key_sentences": instructions.KeySentenceChecker,
81
+ _KEYWORD + "forbidden_words": {_KEYWORD + "forbidden_words"},
82
+ _KEYWORD + "letter_frequency": {_KEYWORD + "letter_frequency"},
83
+ _LANGUAGE + "response_language": {
84
+ _LANGUAGE + "response_language",
85
+ _FORMAT + "multiple_sections",
86
+ _KEYWORD + "existence",
87
+ _KEYWORD + "frequency",
88
+ _KEYWORD + "forbidden_words",
89
+ _STARTEND + "end_checker",
90
+ _CHANGE_CASES + "english_capital",
91
+ _CHANGE_CASES + "english_lowercase",
92
+ },
93
+ _LENGTH + "number_sentences": {_LENGTH + "number_sentences"},
94
+ _LENGTH + "number_paragraphs": {
95
+ _LENGTH + "number_paragraphs",
96
+ _LENGTH + "nth_paragraph_first_word",
97
+ _LENGTH + "number_sentences",
98
+ _LENGTH + "nth_paragraph_first_word",
99
+ },
100
+ _LENGTH + "number_words": {_LENGTH + "number_words"},
101
+ _LENGTH + "nth_paragraph_first_word": {
102
+ _LENGTH + "nth_paragraph_first_word",
103
+ _LENGTH + "number_paragraphs",
104
+ },
105
+ _CONTENT + "number_placeholders": {_CONTENT + "number_placeholders"},
106
+ _CONTENT + "postscript": {_CONTENT + "postscript"},
107
+ _FORMAT + "number_bullet_lists": {_FORMAT + "number_bullet_lists"},
108
+ # TODO(jeffreyzhou): Pre-create paragraph or use prompt to replace
109
+ # _CONTENT + "rephrase_paragraph": instructions.RephraseParagraph,
110
+ _FORMAT + "constrained_response": set(INSTRUCTION_DICT.keys()),
111
+ _FORMAT + "number_highlighted_sections": {_FORMAT + "number_highlighted_sections"},
112
+ _FORMAT + "multiple_sections": {
113
+ _FORMAT + "multiple_sections",
114
+ _LANGUAGE + "response_language",
115
+ _FORMAT + "number_highlighted_sections",
116
+ },
117
+ # TODO(tianjianlu): Re-enable rephrasing with preprocessing the message.
118
+ # _FORMAT + "rephrase": instructions.RephraseChecker,
119
+ _FORMAT + "json_format": set(INSTRUCTION_DICT.keys()).difference(
120
+ {_KEYWORD + "forbidden_words", _KEYWORD + "existence"}
121
+ ),
122
+ _FORMAT + "title": {_FORMAT + "title"},
123
+ # TODO(tianjianlu): Re-enable with specific prompts.
124
+ # _MULTITURN + "constrained_start": instructions.ConstrainedStartChecker,
125
+ _COMBINATION + "two_responses": set(INSTRUCTION_DICT.keys()).difference(
126
+ {
127
+ _KEYWORD + "forbidden_words",
128
+ _KEYWORD + "existence",
129
+ _LANGUAGE + "response_language",
130
+ _FORMAT + "title",
131
+ _PUNCTUATION + "no_comma",
132
+ }
133
+ ),
134
+ _COMBINATION + "repeat_prompt": set(INSTRUCTION_DICT.keys()).difference(
135
+ {_KEYWORD + "existence", _FORMAT + "title", _PUNCTUATION + "no_comma"}
136
+ ),
137
+ _STARTEND + "end_checker": {_STARTEND + "end_checker"},
138
+ _CHANGE_CASES + "capital_word_frequency": {
139
+ _CHANGE_CASES + "capital_word_frequency",
140
+ _CHANGE_CASES + "english_lowercase",
141
+ _CHANGE_CASES + "english_capital",
142
+ },
143
+ _CHANGE_CASES + "english_capital": {_CHANGE_CASES + "english_capital"},
144
+ _CHANGE_CASES + "english_lowercase": {
145
+ _CHANGE_CASES + "english_lowercase",
146
+ _CHANGE_CASES + "english_capital",
147
+ },
148
+ _PUNCTUATION + "no_comma": {_PUNCTUATION + "no_comma"},
149
+ _STARTEND + "quotation": {_STARTEND + "quotation", _FORMAT + "title"},
150
+ }
151
+
152
+
153
+ def conflict_make(conflicts):
154
+ """Makes sure if A conflicts with B, B will conflict with A.
155
+
156
+ Args:
157
+ conflicts: Dictionary of potential conflicts where key is instruction id
158
+ and value is set of instruction ids that it conflicts with.
159
+
160
+ Returns:
161
+ Revised version of the dictionary. All instructions conflict with
162
+ themselves. If A conflicts with B, B will conflict with A.
163
+ """
164
+ for key in conflicts:
165
+ for k in conflicts[key]:
166
+ conflicts[k].add(key)
167
+ conflicts[key].add(key)
168
+ return conflicts
eval/ifeval/instructions_util.py ADDED
@@ -0,0 +1,1701 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 The Google Research Authors.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ """Utility library of instructions."""
16
+
17
+ import functools
18
+ import os
19
+ import random
20
+ import re
21
+ from importlib.metadata import version
22
+
23
+ import immutabledict
24
+ import nltk
25
+ from packaging.version import parse as parse_version
26
+
27
+
28
+ # Downloading 'punkt' with nltk<3.9 has a remote code vuln.
29
+ # see https://github.com/EleutherAI/lm-evaluation-harness/issues/2210
30
+ # and https://github.com/nltk/nltk/issues/3266
31
+ # for more information.
32
+ NLTK_MIN_VERSION = "3.9.1"
33
+ RANK = os.environ.get("LOCAL_RANK", "0")
34
+
35
+
36
+ def download_nltk_resources():
37
+ """Download 'punkt' if not already installed"""
38
+ assert (nltk_version := parse_version(version("nltk"))) >= parse_version(
39
+ NLTK_MIN_VERSION
40
+ ), (
41
+ f"`nltk` version {nltk_version} is not >= {NLTK_MIN_VERSION}. Please update `nltk` before proceeding--older versions are vulnerable to a remote code execution vulnerability."
42
+ )
43
+
44
+ try:
45
+ nltk.data.find("tokenizers/punkt_tab")
46
+ except LookupError:
47
+ if RANK == "0":
48
+ nltk.download("punkt_tab")
49
+ print("Downloaded punkt_tab on rank 0")
50
+
51
+
52
+ download_nltk_resources()
53
+
54
+ WORD_LIST = [
55
+ "western",
56
+ "sentence",
57
+ "signal",
58
+ "dump",
59
+ "spot",
60
+ "opposite",
61
+ "bottom",
62
+ "potato",
63
+ "administration",
64
+ "working",
65
+ "welcome",
66
+ "morning",
67
+ "good",
68
+ "agency",
69
+ "primary",
70
+ "wish",
71
+ "responsibility",
72
+ "press",
73
+ "problem",
74
+ "president",
75
+ "steal",
76
+ "brush",
77
+ "read",
78
+ "type",
79
+ "beat",
80
+ "trainer",
81
+ "growth",
82
+ "lock",
83
+ "bone",
84
+ "case",
85
+ "equal",
86
+ "comfortable",
87
+ "region",
88
+ "replacement",
89
+ "performance",
90
+ "mate",
91
+ "walk",
92
+ "medicine",
93
+ "film",
94
+ "thing",
95
+ "rock",
96
+ "tap",
97
+ "total",
98
+ "competition",
99
+ "ease",
100
+ "south",
101
+ "establishment",
102
+ "gather",
103
+ "parking",
104
+ "world",
105
+ "plenty",
106
+ "breath",
107
+ "claim",
108
+ "alcohol",
109
+ "trade",
110
+ "dear",
111
+ "highlight",
112
+ "street",
113
+ "matter",
114
+ "decision",
115
+ "mess",
116
+ "agreement",
117
+ "studio",
118
+ "coach",
119
+ "assist",
120
+ "brain",
121
+ "wing",
122
+ "style",
123
+ "private",
124
+ "top",
125
+ "brown",
126
+ "leg",
127
+ "buy",
128
+ "procedure",
129
+ "method",
130
+ "speed",
131
+ "high",
132
+ "company",
133
+ "valuable",
134
+ "pie",
135
+ "analyst",
136
+ "session",
137
+ "pattern",
138
+ "district",
139
+ "pleasure",
140
+ "dinner",
141
+ "swimming",
142
+ "joke",
143
+ "order",
144
+ "plate",
145
+ "department",
146
+ "motor",
147
+ "cell",
148
+ "spend",
149
+ "cabinet",
150
+ "difference",
151
+ "power",
152
+ "examination",
153
+ "engine",
154
+ "horse",
155
+ "dimension",
156
+ "pay",
157
+ "toe",
158
+ "curve",
159
+ "literature",
160
+ "bother",
161
+ "fire",
162
+ "possibility",
163
+ "debate",
164
+ "activity",
165
+ "passage",
166
+ "hello",
167
+ "cycle",
168
+ "background",
169
+ "quiet",
170
+ "author",
171
+ "effect",
172
+ "actor",
173
+ "page",
174
+ "bicycle",
175
+ "error",
176
+ "throat",
177
+ "attack",
178
+ "character",
179
+ "phone",
180
+ "tea",
181
+ "increase",
182
+ "outcome",
183
+ "file",
184
+ "specific",
185
+ "inspector",
186
+ "internal",
187
+ "potential",
188
+ "staff",
189
+ "building",
190
+ "employer",
191
+ "shoe",
192
+ "hand",
193
+ "direction",
194
+ "garden",
195
+ "purchase",
196
+ "interview",
197
+ "study",
198
+ "recognition",
199
+ "member",
200
+ "spiritual",
201
+ "oven",
202
+ "sandwich",
203
+ "weird",
204
+ "passenger",
205
+ "particular",
206
+ "response",
207
+ "reaction",
208
+ "size",
209
+ "variation",
210
+ "a",
211
+ "cancel",
212
+ "candy",
213
+ "exit",
214
+ "guest",
215
+ "condition",
216
+ "fly",
217
+ "price",
218
+ "weakness",
219
+ "convert",
220
+ "hotel",
221
+ "great",
222
+ "mouth",
223
+ "mind",
224
+ "song",
225
+ "sugar",
226
+ "suspect",
227
+ "telephone",
228
+ "ear",
229
+ "roof",
230
+ "paint",
231
+ "refrigerator",
232
+ "organization",
233
+ "jury",
234
+ "reward",
235
+ "engineering",
236
+ "day",
237
+ "possession",
238
+ "crew",
239
+ "bar",
240
+ "road",
241
+ "description",
242
+ "celebration",
243
+ "score",
244
+ "mark",
245
+ "letter",
246
+ "shower",
247
+ "suggestion",
248
+ "sir",
249
+ "luck",
250
+ "national",
251
+ "progress",
252
+ "hall",
253
+ "stroke",
254
+ "theory",
255
+ "offer",
256
+ "story",
257
+ "tax",
258
+ "definition",
259
+ "history",
260
+ "ride",
261
+ "medium",
262
+ "opening",
263
+ "glass",
264
+ "elevator",
265
+ "stomach",
266
+ "question",
267
+ "ability",
268
+ "leading",
269
+ "village",
270
+ "computer",
271
+ "city",
272
+ "grand",
273
+ "confidence",
274
+ "candle",
275
+ "priest",
276
+ "recommendation",
277
+ "point",
278
+ "necessary",
279
+ "body",
280
+ "desk",
281
+ "secret",
282
+ "horror",
283
+ "noise",
284
+ "culture",
285
+ "warning",
286
+ "water",
287
+ "round",
288
+ "diet",
289
+ "flower",
290
+ "bus",
291
+ "tough",
292
+ "permission",
293
+ "week",
294
+ "prompt",
295
+ "connection",
296
+ "abuse",
297
+ "height",
298
+ "save",
299
+ "corner",
300
+ "border",
301
+ "stress",
302
+ "drive",
303
+ "stop",
304
+ "rip",
305
+ "meal",
306
+ "listen",
307
+ "confusion",
308
+ "girlfriend",
309
+ "living",
310
+ "relation",
311
+ "significance",
312
+ "plan",
313
+ "creative",
314
+ "atmosphere",
315
+ "blame",
316
+ "invite",
317
+ "housing",
318
+ "paper",
319
+ "drink",
320
+ "roll",
321
+ "silver",
322
+ "drunk",
323
+ "age",
324
+ "damage",
325
+ "smoke",
326
+ "environment",
327
+ "pack",
328
+ "savings",
329
+ "influence",
330
+ "tourist",
331
+ "rain",
332
+ "post",
333
+ "sign",
334
+ "grandmother",
335
+ "run",
336
+ "profit",
337
+ "push",
338
+ "clerk",
339
+ "final",
340
+ "wine",
341
+ "swim",
342
+ "pause",
343
+ "stuff",
344
+ "singer",
345
+ "funeral",
346
+ "average",
347
+ "source",
348
+ "scene",
349
+ "tradition",
350
+ "personal",
351
+ "snow",
352
+ "nobody",
353
+ "distance",
354
+ "sort",
355
+ "sensitive",
356
+ "animal",
357
+ "major",
358
+ "negotiation",
359
+ "click",
360
+ "mood",
361
+ "period",
362
+ "arrival",
363
+ "expression",
364
+ "holiday",
365
+ "repeat",
366
+ "dust",
367
+ "closet",
368
+ "gold",
369
+ "bad",
370
+ "sail",
371
+ "combination",
372
+ "clothes",
373
+ "emphasis",
374
+ "duty",
375
+ "black",
376
+ "step",
377
+ "school",
378
+ "jump",
379
+ "document",
380
+ "professional",
381
+ "lip",
382
+ "chemical",
383
+ "front",
384
+ "wake",
385
+ "while",
386
+ "inside",
387
+ "watch",
388
+ "row",
389
+ "subject",
390
+ "penalty",
391
+ "balance",
392
+ "possible",
393
+ "adult",
394
+ "aside",
395
+ "sample",
396
+ "appeal",
397
+ "wedding",
398
+ "depth",
399
+ "king",
400
+ "award",
401
+ "wife",
402
+ "blow",
403
+ "site",
404
+ "camp",
405
+ "music",
406
+ "safe",
407
+ "gift",
408
+ "fault",
409
+ "guess",
410
+ "act",
411
+ "shame",
412
+ "drama",
413
+ "capital",
414
+ "exam",
415
+ "stupid",
416
+ "record",
417
+ "sound",
418
+ "swing",
419
+ "novel",
420
+ "minimum",
421
+ "ratio",
422
+ "machine",
423
+ "shape",
424
+ "lead",
425
+ "operation",
426
+ "salary",
427
+ "cloud",
428
+ "affair",
429
+ "hit",
430
+ "chapter",
431
+ "stage",
432
+ "quantity",
433
+ "access",
434
+ "army",
435
+ "chain",
436
+ "traffic",
437
+ "kick",
438
+ "analysis",
439
+ "airport",
440
+ "time",
441
+ "vacation",
442
+ "philosophy",
443
+ "ball",
444
+ "chest",
445
+ "thanks",
446
+ "place",
447
+ "mountain",
448
+ "advertising",
449
+ "red",
450
+ "past",
451
+ "rent",
452
+ "return",
453
+ "tour",
454
+ "house",
455
+ "construction",
456
+ "net",
457
+ "native",
458
+ "war",
459
+ "figure",
460
+ "fee",
461
+ "spray",
462
+ "user",
463
+ "dirt",
464
+ "shot",
465
+ "task",
466
+ "stick",
467
+ "friend",
468
+ "software",
469
+ "promotion",
470
+ "interaction",
471
+ "surround",
472
+ "block",
473
+ "purpose",
474
+ "practice",
475
+ "conflict",
476
+ "routine",
477
+ "requirement",
478
+ "bonus",
479
+ "hole",
480
+ "state",
481
+ "junior",
482
+ "sweet",
483
+ "catch",
484
+ "tear",
485
+ "fold",
486
+ "wall",
487
+ "editor",
488
+ "life",
489
+ "position",
490
+ "pound",
491
+ "respect",
492
+ "bathroom",
493
+ "coat",
494
+ "script",
495
+ "job",
496
+ "teach",
497
+ "birth",
498
+ "view",
499
+ "resolve",
500
+ "theme",
501
+ "employee",
502
+ "doubt",
503
+ "market",
504
+ "education",
505
+ "serve",
506
+ "recover",
507
+ "tone",
508
+ "harm",
509
+ "miss",
510
+ "union",
511
+ "understanding",
512
+ "cow",
513
+ "river",
514
+ "association",
515
+ "concept",
516
+ "training",
517
+ "recipe",
518
+ "relationship",
519
+ "reserve",
520
+ "depression",
521
+ "proof",
522
+ "hair",
523
+ "revenue",
524
+ "independent",
525
+ "lift",
526
+ "assignment",
527
+ "temporary",
528
+ "amount",
529
+ "loss",
530
+ "edge",
531
+ "track",
532
+ "check",
533
+ "rope",
534
+ "estimate",
535
+ "pollution",
536
+ "stable",
537
+ "message",
538
+ "delivery",
539
+ "perspective",
540
+ "mirror",
541
+ "assistant",
542
+ "representative",
543
+ "witness",
544
+ "nature",
545
+ "judge",
546
+ "fruit",
547
+ "tip",
548
+ "devil",
549
+ "town",
550
+ "emergency",
551
+ "upper",
552
+ "drop",
553
+ "stay",
554
+ "human",
555
+ "neck",
556
+ "speaker",
557
+ "network",
558
+ "sing",
559
+ "resist",
560
+ "league",
561
+ "trip",
562
+ "signature",
563
+ "lawyer",
564
+ "importance",
565
+ "gas",
566
+ "choice",
567
+ "engineer",
568
+ "success",
569
+ "part",
570
+ "external",
571
+ "worker",
572
+ "simple",
573
+ "quarter",
574
+ "student",
575
+ "heart",
576
+ "pass",
577
+ "spite",
578
+ "shift",
579
+ "rough",
580
+ "lady",
581
+ "grass",
582
+ "community",
583
+ "garage",
584
+ "youth",
585
+ "standard",
586
+ "skirt",
587
+ "promise",
588
+ "blind",
589
+ "television",
590
+ "disease",
591
+ "commission",
592
+ "positive",
593
+ "energy",
594
+ "calm",
595
+ "presence",
596
+ "tune",
597
+ "basis",
598
+ "preference",
599
+ "head",
600
+ "common",
601
+ "cut",
602
+ "somewhere",
603
+ "presentation",
604
+ "current",
605
+ "thought",
606
+ "revolution",
607
+ "effort",
608
+ "master",
609
+ "implement",
610
+ "republic",
611
+ "floor",
612
+ "principle",
613
+ "stranger",
614
+ "shoulder",
615
+ "grade",
616
+ "button",
617
+ "tennis",
618
+ "police",
619
+ "collection",
620
+ "account",
621
+ "register",
622
+ "glove",
623
+ "divide",
624
+ "professor",
625
+ "chair",
626
+ "priority",
627
+ "combine",
628
+ "peace",
629
+ "extension",
630
+ "maybe",
631
+ "evening",
632
+ "frame",
633
+ "sister",
634
+ "wave",
635
+ "code",
636
+ "application",
637
+ "mouse",
638
+ "match",
639
+ "counter",
640
+ "bottle",
641
+ "half",
642
+ "cheek",
643
+ "resolution",
644
+ "back",
645
+ "knowledge",
646
+ "make",
647
+ "discussion",
648
+ "screw",
649
+ "length",
650
+ "accident",
651
+ "battle",
652
+ "dress",
653
+ "knee",
654
+ "log",
655
+ "package",
656
+ "it",
657
+ "turn",
658
+ "hearing",
659
+ "newspaper",
660
+ "layer",
661
+ "wealth",
662
+ "profile",
663
+ "imagination",
664
+ "answer",
665
+ "weekend",
666
+ "teacher",
667
+ "appearance",
668
+ "meet",
669
+ "bike",
670
+ "rise",
671
+ "belt",
672
+ "crash",
673
+ "bowl",
674
+ "equivalent",
675
+ "support",
676
+ "image",
677
+ "poem",
678
+ "risk",
679
+ "excitement",
680
+ "remote",
681
+ "secretary",
682
+ "public",
683
+ "produce",
684
+ "plane",
685
+ "display",
686
+ "money",
687
+ "sand",
688
+ "situation",
689
+ "punch",
690
+ "customer",
691
+ "title",
692
+ "shake",
693
+ "mortgage",
694
+ "option",
695
+ "number",
696
+ "pop",
697
+ "window",
698
+ "extent",
699
+ "nothing",
700
+ "experience",
701
+ "opinion",
702
+ "departure",
703
+ "dance",
704
+ "indication",
705
+ "boy",
706
+ "material",
707
+ "band",
708
+ "leader",
709
+ "sun",
710
+ "beautiful",
711
+ "muscle",
712
+ "farmer",
713
+ "variety",
714
+ "fat",
715
+ "handle",
716
+ "director",
717
+ "opportunity",
718
+ "calendar",
719
+ "outside",
720
+ "pace",
721
+ "bath",
722
+ "fish",
723
+ "consequence",
724
+ "put",
725
+ "owner",
726
+ "go",
727
+ "doctor",
728
+ "information",
729
+ "share",
730
+ "hurt",
731
+ "protection",
732
+ "career",
733
+ "finance",
734
+ "force",
735
+ "golf",
736
+ "garbage",
737
+ "aspect",
738
+ "kid",
739
+ "food",
740
+ "boot",
741
+ "milk",
742
+ "respond",
743
+ "objective",
744
+ "reality",
745
+ "raw",
746
+ "ring",
747
+ "mall",
748
+ "one",
749
+ "impact",
750
+ "area",
751
+ "news",
752
+ "international",
753
+ "series",
754
+ "impress",
755
+ "mother",
756
+ "shelter",
757
+ "strike",
758
+ "loan",
759
+ "month",
760
+ "seat",
761
+ "anything",
762
+ "entertainment",
763
+ "familiar",
764
+ "clue",
765
+ "year",
766
+ "glad",
767
+ "supermarket",
768
+ "natural",
769
+ "god",
770
+ "cost",
771
+ "conversation",
772
+ "tie",
773
+ "ruin",
774
+ "comfort",
775
+ "earth",
776
+ "storm",
777
+ "percentage",
778
+ "assistance",
779
+ "budget",
780
+ "strength",
781
+ "beginning",
782
+ "sleep",
783
+ "other",
784
+ "young",
785
+ "unit",
786
+ "fill",
787
+ "store",
788
+ "desire",
789
+ "hide",
790
+ "value",
791
+ "cup",
792
+ "maintenance",
793
+ "nurse",
794
+ "function",
795
+ "tower",
796
+ "role",
797
+ "class",
798
+ "camera",
799
+ "database",
800
+ "panic",
801
+ "nation",
802
+ "basket",
803
+ "ice",
804
+ "art",
805
+ "spirit",
806
+ "chart",
807
+ "exchange",
808
+ "feedback",
809
+ "statement",
810
+ "reputation",
811
+ "search",
812
+ "hunt",
813
+ "exercise",
814
+ "nasty",
815
+ "notice",
816
+ "male",
817
+ "yard",
818
+ "annual",
819
+ "collar",
820
+ "date",
821
+ "platform",
822
+ "plant",
823
+ "fortune",
824
+ "passion",
825
+ "friendship",
826
+ "spread",
827
+ "cancer",
828
+ "ticket",
829
+ "attitude",
830
+ "island",
831
+ "active",
832
+ "object",
833
+ "service",
834
+ "buyer",
835
+ "bite",
836
+ "card",
837
+ "face",
838
+ "steak",
839
+ "proposal",
840
+ "patient",
841
+ "heat",
842
+ "rule",
843
+ "resident",
844
+ "broad",
845
+ "politics",
846
+ "west",
847
+ "knife",
848
+ "expert",
849
+ "girl",
850
+ "design",
851
+ "salt",
852
+ "baseball",
853
+ "grab",
854
+ "inspection",
855
+ "cousin",
856
+ "couple",
857
+ "magazine",
858
+ "cook",
859
+ "dependent",
860
+ "security",
861
+ "chicken",
862
+ "version",
863
+ "currency",
864
+ "ladder",
865
+ "scheme",
866
+ "kitchen",
867
+ "employment",
868
+ "local",
869
+ "attention",
870
+ "manager",
871
+ "fact",
872
+ "cover",
873
+ "sad",
874
+ "guard",
875
+ "relative",
876
+ "county",
877
+ "rate",
878
+ "lunch",
879
+ "program",
880
+ "initiative",
881
+ "gear",
882
+ "bridge",
883
+ "breast",
884
+ "talk",
885
+ "dish",
886
+ "guarantee",
887
+ "beer",
888
+ "vehicle",
889
+ "reception",
890
+ "woman",
891
+ "substance",
892
+ "copy",
893
+ "lecture",
894
+ "advantage",
895
+ "park",
896
+ "cold",
897
+ "death",
898
+ "mix",
899
+ "hold",
900
+ "scale",
901
+ "tomorrow",
902
+ "blood",
903
+ "request",
904
+ "green",
905
+ "cookie",
906
+ "church",
907
+ "strip",
908
+ "forever",
909
+ "beyond",
910
+ "debt",
911
+ "tackle",
912
+ "wash",
913
+ "following",
914
+ "feel",
915
+ "maximum",
916
+ "sector",
917
+ "sea",
918
+ "property",
919
+ "economics",
920
+ "menu",
921
+ "bench",
922
+ "try",
923
+ "language",
924
+ "start",
925
+ "call",
926
+ "solid",
927
+ "address",
928
+ "income",
929
+ "foot",
930
+ "senior",
931
+ "honey",
932
+ "few",
933
+ "mixture",
934
+ "cash",
935
+ "grocery",
936
+ "link",
937
+ "map",
938
+ "form",
939
+ "factor",
940
+ "pot",
941
+ "model",
942
+ "writer",
943
+ "farm",
944
+ "winter",
945
+ "skill",
946
+ "anywhere",
947
+ "birthday",
948
+ "policy",
949
+ "release",
950
+ "husband",
951
+ "lab",
952
+ "hurry",
953
+ "mail",
954
+ "equipment",
955
+ "sink",
956
+ "pair",
957
+ "driver",
958
+ "consideration",
959
+ "leather",
960
+ "skin",
961
+ "blue",
962
+ "boat",
963
+ "sale",
964
+ "brick",
965
+ "two",
966
+ "feed",
967
+ "square",
968
+ "dot",
969
+ "rush",
970
+ "dream",
971
+ "location",
972
+ "afternoon",
973
+ "manufacturer",
974
+ "control",
975
+ "occasion",
976
+ "trouble",
977
+ "introduction",
978
+ "advice",
979
+ "bet",
980
+ "eat",
981
+ "kill",
982
+ "category",
983
+ "manner",
984
+ "office",
985
+ "estate",
986
+ "pride",
987
+ "awareness",
988
+ "slip",
989
+ "crack",
990
+ "client",
991
+ "nail",
992
+ "shoot",
993
+ "membership",
994
+ "soft",
995
+ "anybody",
996
+ "web",
997
+ "official",
998
+ "individual",
999
+ "pizza",
1000
+ "interest",
1001
+ "bag",
1002
+ "spell",
1003
+ "profession",
1004
+ "queen",
1005
+ "deal",
1006
+ "resource",
1007
+ "ship",
1008
+ "guy",
1009
+ "chocolate",
1010
+ "joint",
1011
+ "formal",
1012
+ "upstairs",
1013
+ "car",
1014
+ "resort",
1015
+ "abroad",
1016
+ "dealer",
1017
+ "associate",
1018
+ "finger",
1019
+ "surgery",
1020
+ "comment",
1021
+ "team",
1022
+ "detail",
1023
+ "crazy",
1024
+ "path",
1025
+ "tale",
1026
+ "initial",
1027
+ "arm",
1028
+ "radio",
1029
+ "demand",
1030
+ "single",
1031
+ "draw",
1032
+ "yellow",
1033
+ "contest",
1034
+ "piece",
1035
+ "quote",
1036
+ "pull",
1037
+ "commercial",
1038
+ "shirt",
1039
+ "contribution",
1040
+ "cream",
1041
+ "channel",
1042
+ "suit",
1043
+ "discipline",
1044
+ "instruction",
1045
+ "concert",
1046
+ "speech",
1047
+ "low",
1048
+ "effective",
1049
+ "hang",
1050
+ "scratch",
1051
+ "industry",
1052
+ "breakfast",
1053
+ "lay",
1054
+ "join",
1055
+ "metal",
1056
+ "bedroom",
1057
+ "minute",
1058
+ "product",
1059
+ "rest",
1060
+ "temperature",
1061
+ "many",
1062
+ "give",
1063
+ "argument",
1064
+ "print",
1065
+ "purple",
1066
+ "laugh",
1067
+ "health",
1068
+ "credit",
1069
+ "investment",
1070
+ "sell",
1071
+ "setting",
1072
+ "lesson",
1073
+ "egg",
1074
+ "middle",
1075
+ "marriage",
1076
+ "level",
1077
+ "evidence",
1078
+ "phrase",
1079
+ "love",
1080
+ "self",
1081
+ "benefit",
1082
+ "guidance",
1083
+ "affect",
1084
+ "you",
1085
+ "dad",
1086
+ "anxiety",
1087
+ "special",
1088
+ "boyfriend",
1089
+ "test",
1090
+ "blank",
1091
+ "payment",
1092
+ "soup",
1093
+ "obligation",
1094
+ "reply",
1095
+ "smile",
1096
+ "deep",
1097
+ "complaint",
1098
+ "addition",
1099
+ "review",
1100
+ "box",
1101
+ "towel",
1102
+ "minor",
1103
+ "fun",
1104
+ "soil",
1105
+ "issue",
1106
+ "cigarette",
1107
+ "internet",
1108
+ "gain",
1109
+ "tell",
1110
+ "entry",
1111
+ "spare",
1112
+ "incident",
1113
+ "family",
1114
+ "refuse",
1115
+ "branch",
1116
+ "can",
1117
+ "pen",
1118
+ "grandfather",
1119
+ "constant",
1120
+ "tank",
1121
+ "uncle",
1122
+ "climate",
1123
+ "ground",
1124
+ "volume",
1125
+ "communication",
1126
+ "kind",
1127
+ "poet",
1128
+ "child",
1129
+ "screen",
1130
+ "mine",
1131
+ "quit",
1132
+ "gene",
1133
+ "lack",
1134
+ "charity",
1135
+ "memory",
1136
+ "tooth",
1137
+ "fear",
1138
+ "mention",
1139
+ "marketing",
1140
+ "reveal",
1141
+ "reason",
1142
+ "court",
1143
+ "season",
1144
+ "freedom",
1145
+ "land",
1146
+ "sport",
1147
+ "audience",
1148
+ "classroom",
1149
+ "law",
1150
+ "hook",
1151
+ "win",
1152
+ "carry",
1153
+ "eye",
1154
+ "smell",
1155
+ "distribution",
1156
+ "research",
1157
+ "country",
1158
+ "dare",
1159
+ "hope",
1160
+ "whereas",
1161
+ "stretch",
1162
+ "library",
1163
+ "if",
1164
+ "delay",
1165
+ "college",
1166
+ "plastic",
1167
+ "book",
1168
+ "present",
1169
+ "use",
1170
+ "worry",
1171
+ "champion",
1172
+ "goal",
1173
+ "economy",
1174
+ "march",
1175
+ "election",
1176
+ "reflection",
1177
+ "midnight",
1178
+ "slide",
1179
+ "inflation",
1180
+ "action",
1181
+ "challenge",
1182
+ "guitar",
1183
+ "coast",
1184
+ "apple",
1185
+ "campaign",
1186
+ "field",
1187
+ "jacket",
1188
+ "sense",
1189
+ "way",
1190
+ "visual",
1191
+ "remove",
1192
+ "weather",
1193
+ "trash",
1194
+ "cable",
1195
+ "regret",
1196
+ "buddy",
1197
+ "beach",
1198
+ "historian",
1199
+ "courage",
1200
+ "sympathy",
1201
+ "truck",
1202
+ "tension",
1203
+ "permit",
1204
+ "nose",
1205
+ "bed",
1206
+ "son",
1207
+ "person",
1208
+ "base",
1209
+ "meat",
1210
+ "usual",
1211
+ "air",
1212
+ "meeting",
1213
+ "worth",
1214
+ "game",
1215
+ "independence",
1216
+ "physical",
1217
+ "brief",
1218
+ "play",
1219
+ "raise",
1220
+ "board",
1221
+ "she",
1222
+ "key",
1223
+ "writing",
1224
+ "pick",
1225
+ "command",
1226
+ "party",
1227
+ "yesterday",
1228
+ "spring",
1229
+ "candidate",
1230
+ "physics",
1231
+ "university",
1232
+ "concern",
1233
+ "development",
1234
+ "change",
1235
+ "string",
1236
+ "target",
1237
+ "instance",
1238
+ "room",
1239
+ "bitter",
1240
+ "bird",
1241
+ "football",
1242
+ "normal",
1243
+ "split",
1244
+ "impression",
1245
+ "wood",
1246
+ "long",
1247
+ "meaning",
1248
+ "stock",
1249
+ "cap",
1250
+ "leadership",
1251
+ "media",
1252
+ "ambition",
1253
+ "fishing",
1254
+ "essay",
1255
+ "salad",
1256
+ "repair",
1257
+ "today",
1258
+ "designer",
1259
+ "night",
1260
+ "bank",
1261
+ "drawing",
1262
+ "inevitable",
1263
+ "phase",
1264
+ "vast",
1265
+ "chip",
1266
+ "anger",
1267
+ "switch",
1268
+ "cry",
1269
+ "twist",
1270
+ "personality",
1271
+ "attempt",
1272
+ "storage",
1273
+ "being",
1274
+ "preparation",
1275
+ "bat",
1276
+ "selection",
1277
+ "white",
1278
+ "technology",
1279
+ "contract",
1280
+ "side",
1281
+ "section",
1282
+ "station",
1283
+ "till",
1284
+ "structure",
1285
+ "tongue",
1286
+ "taste",
1287
+ "truth",
1288
+ "difficulty",
1289
+ "group",
1290
+ "limit",
1291
+ "main",
1292
+ "move",
1293
+ "feeling",
1294
+ "light",
1295
+ "example",
1296
+ "mission",
1297
+ "might",
1298
+ "wait",
1299
+ "wheel",
1300
+ "shop",
1301
+ "host",
1302
+ "classic",
1303
+ "alternative",
1304
+ "cause",
1305
+ "agent",
1306
+ "consist",
1307
+ "table",
1308
+ "airline",
1309
+ "text",
1310
+ "pool",
1311
+ "craft",
1312
+ "range",
1313
+ "fuel",
1314
+ "tool",
1315
+ "partner",
1316
+ "load",
1317
+ "entrance",
1318
+ "deposit",
1319
+ "hate",
1320
+ "article",
1321
+ "video",
1322
+ "summer",
1323
+ "feature",
1324
+ "extreme",
1325
+ "mobile",
1326
+ "hospital",
1327
+ "flight",
1328
+ "fall",
1329
+ "pension",
1330
+ "piano",
1331
+ "fail",
1332
+ "result",
1333
+ "rub",
1334
+ "gap",
1335
+ "system",
1336
+ "report",
1337
+ "suck",
1338
+ "ordinary",
1339
+ "wind",
1340
+ "nerve",
1341
+ "ask",
1342
+ "shine",
1343
+ "note",
1344
+ "line",
1345
+ "mom",
1346
+ "perception",
1347
+ "brother",
1348
+ "reference",
1349
+ "bend",
1350
+ "charge",
1351
+ "treat",
1352
+ "trick",
1353
+ "term",
1354
+ "homework",
1355
+ "bake",
1356
+ "bid",
1357
+ "status",
1358
+ "project",
1359
+ "strategy",
1360
+ "orange",
1361
+ "let",
1362
+ "enthusiasm",
1363
+ "parent",
1364
+ "concentrate",
1365
+ "device",
1366
+ "travel",
1367
+ "poetry",
1368
+ "business",
1369
+ "society",
1370
+ "kiss",
1371
+ "end",
1372
+ "vegetable",
1373
+ "employ",
1374
+ "schedule",
1375
+ "hour",
1376
+ "brave",
1377
+ "focus",
1378
+ "process",
1379
+ "movie",
1380
+ "illegal",
1381
+ "general",
1382
+ "coffee",
1383
+ "ad",
1384
+ "highway",
1385
+ "chemistry",
1386
+ "psychology",
1387
+ "hire",
1388
+ "bell",
1389
+ "conference",
1390
+ "relief",
1391
+ "show",
1392
+ "neat",
1393
+ "funny",
1394
+ "weight",
1395
+ "quality",
1396
+ "club",
1397
+ "daughter",
1398
+ "zone",
1399
+ "touch",
1400
+ "tonight",
1401
+ "shock",
1402
+ "burn",
1403
+ "excuse",
1404
+ "name",
1405
+ "survey",
1406
+ "landscape",
1407
+ "advance",
1408
+ "satisfaction",
1409
+ "bread",
1410
+ "disaster",
1411
+ "item",
1412
+ "hat",
1413
+ "prior",
1414
+ "shopping",
1415
+ "visit",
1416
+ "east",
1417
+ "photo",
1418
+ "home",
1419
+ "idea",
1420
+ "father",
1421
+ "comparison",
1422
+ "cat",
1423
+ "pipe",
1424
+ "winner",
1425
+ "count",
1426
+ "lake",
1427
+ "fight",
1428
+ "prize",
1429
+ "foundation",
1430
+ "dog",
1431
+ "keep",
1432
+ "ideal",
1433
+ "fan",
1434
+ "struggle",
1435
+ "peak",
1436
+ "safety",
1437
+ "solution",
1438
+ "hell",
1439
+ "conclusion",
1440
+ "population",
1441
+ "strain",
1442
+ "alarm",
1443
+ "measurement",
1444
+ "second",
1445
+ "train",
1446
+ "race",
1447
+ "due",
1448
+ "insurance",
1449
+ "boss",
1450
+ "tree",
1451
+ "monitor",
1452
+ "sick",
1453
+ "course",
1454
+ "drag",
1455
+ "appointment",
1456
+ "slice",
1457
+ "still",
1458
+ "care",
1459
+ "patience",
1460
+ "rich",
1461
+ "escape",
1462
+ "emotion",
1463
+ "royal",
1464
+ "female",
1465
+ "childhood",
1466
+ "government",
1467
+ "picture",
1468
+ "will",
1469
+ "sock",
1470
+ "big",
1471
+ "gate",
1472
+ "oil",
1473
+ "cross",
1474
+ "pin",
1475
+ "improvement",
1476
+ "championship",
1477
+ "silly",
1478
+ "help",
1479
+ "sky",
1480
+ "pitch",
1481
+ "man",
1482
+ "diamond",
1483
+ "most",
1484
+ "transition",
1485
+ "work",
1486
+ "science",
1487
+ "committee",
1488
+ "moment",
1489
+ "fix",
1490
+ "teaching",
1491
+ "dig",
1492
+ "specialist",
1493
+ "complex",
1494
+ "guide",
1495
+ "people",
1496
+ "dead",
1497
+ "voice",
1498
+ "original",
1499
+ "break",
1500
+ "topic",
1501
+ "data",
1502
+ "degree",
1503
+ "reading",
1504
+ "recording",
1505
+ "bunch",
1506
+ "reach",
1507
+ "judgment",
1508
+ "lie",
1509
+ "regular",
1510
+ "set",
1511
+ "painting",
1512
+ "mode",
1513
+ "list",
1514
+ "player",
1515
+ "bear",
1516
+ "north",
1517
+ "wonder",
1518
+ "carpet",
1519
+ "heavy",
1520
+ "officer",
1521
+ "negative",
1522
+ "clock",
1523
+ "unique",
1524
+ "baby",
1525
+ "pain",
1526
+ "assumption",
1527
+ "disk",
1528
+ "iron",
1529
+ "bill",
1530
+ "drawer",
1531
+ "look",
1532
+ "double",
1533
+ "mistake",
1534
+ "finish",
1535
+ "future",
1536
+ "brilliant",
1537
+ "contact",
1538
+ "math",
1539
+ "rice",
1540
+ "leave",
1541
+ "restaurant",
1542
+ "discount",
1543
+ "sex",
1544
+ "virus",
1545
+ "bit",
1546
+ "trust",
1547
+ "event",
1548
+ "wear",
1549
+ "juice",
1550
+ "failure",
1551
+ "bug",
1552
+ "context",
1553
+ "mud",
1554
+ "whole",
1555
+ "wrap",
1556
+ "intention",
1557
+ "draft",
1558
+ "pressure",
1559
+ "cake",
1560
+ "dark",
1561
+ "explanation",
1562
+ "space",
1563
+ "angle",
1564
+ "word",
1565
+ "efficiency",
1566
+ "management",
1567
+ "habit",
1568
+ "star",
1569
+ "chance",
1570
+ "finding",
1571
+ "transportation",
1572
+ "stand",
1573
+ "criticism",
1574
+ "flow",
1575
+ "door",
1576
+ "injury",
1577
+ "insect",
1578
+ "surprise",
1579
+ "apartment",
1580
+ ] # pylint: disable=line-too-long
1581
+
1582
+ # ISO 639-1 codes to language names.
1583
+ LANGUAGE_CODES = immutabledict.immutabledict(
1584
+ {
1585
+ "en": "English",
1586
+ "es": "Spanish",
1587
+ "pt": "Portuguese",
1588
+ "ar": "Arabic",
1589
+ "hi": "Hindi",
1590
+ "fr": "French",
1591
+ "ru": "Russian",
1592
+ "de": "German",
1593
+ "ja": "Japanese",
1594
+ "it": "Italian",
1595
+ "bn": "Bengali",
1596
+ "uk": "Ukrainian",
1597
+ "th": "Thai",
1598
+ "ur": "Urdu",
1599
+ "ta": "Tamil",
1600
+ "te": "Telugu",
1601
+ "bg": "Bulgarian",
1602
+ "ko": "Korean",
1603
+ "pl": "Polish",
1604
+ "he": "Hebrew",
1605
+ "fa": "Persian",
1606
+ "vi": "Vietnamese",
1607
+ "ne": "Nepali",
1608
+ "sw": "Swahili",
1609
+ "kn": "Kannada",
1610
+ "mr": "Marathi",
1611
+ "gu": "Gujarati",
1612
+ "pa": "Punjabi",
1613
+ "ml": "Malayalam",
1614
+ "fi": "Finnish",
1615
+ }
1616
+ )
1617
+
1618
+ _ALPHABETS = "([A-Za-z])"
1619
+ _PREFIXES = "(Mr|St|Mrs|Ms|Dr)[.]"
1620
+ _SUFFIXES = "(Inc|Ltd|Jr|Sr|Co)"
1621
+ _STARTERS = r"(Mr|Mrs|Ms|Dr|Prof|Capt|Cpt|Lt|He\s|She\s|It\s|They\s|Their\s|Our\s|We\s|But\s|However\s|That\s|This\s|Wherever)"
1622
+ _ACRONYMS = "([A-Z][.][A-Z][.](?:[A-Z][.])?)"
1623
+ _WEBSITES = "[.](com|net|org|io|gov|edu|me)"
1624
+ _DIGITS = "([0-9])"
1625
+ _MULTIPLE_DOTS = r"\.{2,}"
1626
+
1627
+
1628
+ def split_into_sentences(text):
1629
+ """Split the text into sentences.
1630
+
1631
+ Args:
1632
+ text: A string that consists of more than or equal to one sentences.
1633
+
1634
+ Returns:
1635
+ A list of strings where each string is a sentence.
1636
+ """
1637
+ text = " " + text + " "
1638
+ text = text.replace("\n", " ")
1639
+ text = re.sub(_PREFIXES, "\\1<prd>", text)
1640
+ text = re.sub(_WEBSITES, "<prd>\\1", text)
1641
+ text = re.sub(_DIGITS + "[.]" + _DIGITS, "\\1<prd>\\2", text)
1642
+ text = re.sub(
1643
+ _MULTIPLE_DOTS,
1644
+ lambda match: "<prd>" * len(match.group(0)) + "<stop>",
1645
+ text,
1646
+ )
1647
+ if "Ph.D" in text:
1648
+ text = text.replace("Ph.D.", "Ph<prd>D<prd>")
1649
+ text = re.sub(r"\s" + _ALPHABETS + "[.] ", " \\1<prd> ", text)
1650
+ text = re.sub(_ACRONYMS + " " + _STARTERS, "\\1<stop> \\2", text)
1651
+ text = re.sub(
1652
+ _ALPHABETS + "[.]" + _ALPHABETS + "[.]" + _ALPHABETS + "[.]",
1653
+ "\\1<prd>\\2<prd>\\3<prd>",
1654
+ text,
1655
+ )
1656
+ text = re.sub(_ALPHABETS + "[.]" + _ALPHABETS + "[.]", "\\1<prd>\\2<prd>", text)
1657
+ text = re.sub(" " + _SUFFIXES + "[.] " + _STARTERS, " \\1<stop> \\2", text)
1658
+ text = re.sub(" " + _SUFFIXES + "[.]", " \\1<prd>", text)
1659
+ text = re.sub(" " + _ALPHABETS + "[.]", " \\1<prd>", text)
1660
+ if "”" in text:
1661
+ text = text.replace(".”", "”.")
1662
+ if '"' in text:
1663
+ text = text.replace('."', '".')
1664
+ if "!" in text:
1665
+ text = text.replace('!"', '"!')
1666
+ if "?" in text:
1667
+ text = text.replace('?"', '"?')
1668
+ text = text.replace(".", ".<stop>")
1669
+ text = text.replace("?", "?<stop>")
1670
+ text = text.replace("!", "!<stop>")
1671
+ text = text.replace("<prd>", ".")
1672
+ sentences = text.split("<stop>")
1673
+ sentences = [s.strip() for s in sentences]
1674
+ if sentences and not sentences[-1]:
1675
+ sentences = sentences[:-1]
1676
+ return sentences
1677
+
1678
+
1679
+ def count_words(text):
1680
+ """Counts the number of words."""
1681
+ tokenizer = nltk.tokenize.RegexpTokenizer(r"\w+")
1682
+ tokens = tokenizer.tokenize(text)
1683
+ num_words = len(tokens)
1684
+ return num_words
1685
+
1686
+
1687
+ @functools.lru_cache(maxsize=None)
1688
+ def _get_sentence_tokenizer():
1689
+ return nltk.data.load("nltk:tokenizers/punkt/english.pickle")
1690
+
1691
+
1692
+ def count_sentences(text):
1693
+ """Count the number of sentences."""
1694
+ tokenizer = _get_sentence_tokenizer()
1695
+ tokenized_sentences = tokenizer.tokenize(text)
1696
+ return len(tokenized_sentences)
1697
+
1698
+
1699
+ def generate_keywords(num_keywords):
1700
+ """Randomly generates a few keywords."""
1701
+ return random.sample(WORD_LIST, k=num_keywords)
eval/ifeval/utils.py ADDED
@@ -0,0 +1,134 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import dataclasses
2
+ from typing import Dict, Optional, Union
3
+
4
+ from . import instructions_registry
5
+
6
+
7
+ @dataclasses.dataclass
8
+ class InputExample:
9
+ key: int
10
+ instruction_id_list: list[str]
11
+ prompt: str
12
+ kwargs: list[Dict[str, Optional[Union[str, int]]]]
13
+
14
+
15
+ @dataclasses.dataclass
16
+ class OutputExample:
17
+ instruction_id_list: list[str]
18
+ prompt: str
19
+ response: str
20
+ follow_all_instructions: bool
21
+ follow_instruction_list: list[bool]
22
+
23
+
24
+ def test_instruction_following_strict(
25
+ inp,
26
+ response,
27
+ ):
28
+ """Tests response to see if instructions are followed."""
29
+ instruction_list = inp.instruction_id_list
30
+ is_following_list = []
31
+
32
+ for index, instruction_id in enumerate(instruction_list):
33
+ instruction_cls = instructions_registry.INSTRUCTION_DICT[instruction_id]
34
+ instruction = instruction_cls(instruction_id)
35
+
36
+ # Remove None values from kwargs to avoid unexpected keyword argument errors in build_description method.
37
+ kwargs = {k: v for k, v in inp.kwargs[index].items() if v}
38
+ instruction.build_description(**kwargs)
39
+ args = instruction.get_instruction_args()
40
+ if args and "prompt" in args:
41
+ instruction.build_description(prompt=inp.prompt)
42
+
43
+ if response.strip() and instruction.check_following(response):
44
+ is_following_list.append(True)
45
+ else:
46
+ is_following_list.append(False)
47
+
48
+ return OutputExample(
49
+ instruction_id_list=inp.instruction_id_list,
50
+ prompt=inp.prompt,
51
+ response=response,
52
+ follow_all_instructions=all(is_following_list),
53
+ follow_instruction_list=is_following_list,
54
+ )
55
+
56
+
57
+ def test_instruction_following_loose(
58
+ inp,
59
+ response,
60
+ ):
61
+ """Tests response for an upper bound for following instructions."""
62
+ r = response.split("\n")
63
+ response_remove_first = "\n".join(r[1:]).strip()
64
+ response_remove_last = "\n".join(r[:-1]).strip()
65
+ response_remove_both = "\n".join(r[1:-1]).strip()
66
+ revised_response = response.replace("*", "")
67
+ revised_response_remove_first = response_remove_first.replace("*", "")
68
+ revised_response_remove_last = response_remove_last.replace("*", "")
69
+ revised_response_remove_both = response_remove_both.replace("*", "")
70
+ all_responses = [
71
+ response,
72
+ revised_response,
73
+ response_remove_first,
74
+ response_remove_last,
75
+ response_remove_both,
76
+ revised_response_remove_first,
77
+ revised_response_remove_last,
78
+ revised_response_remove_both,
79
+ ]
80
+ instruction_list = inp.instruction_id_list
81
+ is_following_list = []
82
+
83
+ for index, instruction_id in enumerate(instruction_list):
84
+ instruction_cls = instructions_registry.INSTRUCTION_DICT[instruction_id]
85
+ instruction = instruction_cls(instruction_id)
86
+
87
+ # Remove None values from kwargs to avoid unexpected keyword argument errors in build_description method.
88
+ kwargs = {k: v for k, v in inp.kwargs[index].items() if v}
89
+ instruction.build_description(**kwargs)
90
+ args = instruction.get_instruction_args()
91
+ if args and "prompt" in args:
92
+ instruction.build_description(prompt=inp.prompt)
93
+
94
+ is_following = False
95
+ for r in all_responses:
96
+ if r.strip() and instruction.check_following(r):
97
+ is_following = True
98
+ break
99
+
100
+ is_following_list.append(is_following)
101
+
102
+ return OutputExample(
103
+ instruction_id_list=inp.instruction_id_list,
104
+ prompt=inp.prompt,
105
+ response=response,
106
+ follow_all_instructions=all(is_following_list),
107
+ follow_instruction_list=is_following_list,
108
+ )
109
+
110
+
111
+ def process_results(doc, results):
112
+ inp = InputExample(
113
+ key=doc["key"],
114
+ instruction_id_list=doc["instruction_id_list"],
115
+ prompt=doc["prompt"],
116
+ kwargs=doc["kwargs"],
117
+ )
118
+ response = results[0]
119
+
120
+ out_strict = test_instruction_following_strict(inp, response)
121
+ out_loose = test_instruction_following_loose(inp, response)
122
+
123
+ return {
124
+ "prompt_level_strict_acc": out_strict.follow_all_instructions,
125
+ "inst_level_strict_acc": out_strict.follow_instruction_list,
126
+ "prompt_level_loose_acc": out_loose.follow_all_instructions,
127
+ "inst_level_loose_acc": out_loose.follow_instruction_list,
128
+ }
129
+
130
+
131
+ def agg_inst_level_acc(items):
132
+ flat_items = [item for sublist in items for item in sublist]
133
+ inst_level_acc = sum(flat_items) / len(flat_items)
134
+ return inst_level_acc
eval/job_a.py ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Sentinel S1 Job A — evidence run. The unmodified starting model answers each frozen question 5 times;
2
+ each answer is scored right/wrong by script against the known answer. No training, no model changes."""
3
+ import argparse, hashlib, json, os, re, string, sys, time
4
+ MODEL, REV, SEED, N_TRIES = "allenai/Olmo-3-7B-Instruct", "6e5971d9eba42665f5bd5a0fcf047f299ce1dccc", 20261112, 5
5
+ PROMPT = {"fact": "Answer the question with a short factual answer only (a few words, no explanation).\n\nQuestion: {q}",
6
+ "math": "Solve the problem. Show your working briefly, then give the final answer on the last line in the form 'Answer: <number>'.\n\nProblem: {q}"}
7
+ MAXTOK = {"fact": 48, "math": 640}
8
+
9
+ def norm(s):
10
+ s = s.lower(); s = "".join(c for c in s if c not in set(string.punctuation))
11
+ s = re.sub(r"\b(a|an|the)\b", " ", s); return " ".join(s.split())
12
+
13
+ def fact_ok(pred, golds):
14
+ p = f" {norm(pred)} "
15
+ return any(f" {norm(g)} " in p for g in golds if norm(g))
16
+
17
+ NUM = re.compile(r"-?\d[\d,]*\.?\d*")
18
+ def math_val(text):
19
+ m = re.findall(r"answer\s*[:=]\s*\$?\s*(-?\d[\d,]*\.?\d*)", text, flags=re.I)
20
+ cand = m[-1] if m else (NUM.findall(text)[-1] if NUM.findall(text) else None)
21
+ if cand is None: return None
22
+ try: return float(cand.replace(",", "").rstrip("."))
23
+ except ValueError: return None
24
+
25
+ def math_ok(pred, golds):
26
+ v = math_val(pred)
27
+ try: g = float(golds[0])
28
+ except ValueError: return False
29
+ return v is not None and abs(v - g) < 1e-6
30
+
31
+ def main():
32
+ ap = argparse.ArgumentParser(); ap.add_argument("--questions"); ap.add_argument("--out"); ap.add_argument("--limit", type=int, default=0); a = ap.parse_args()
33
+ raw = open(a.questions, "rb").read(); qsha = hashlib.sha256(raw).hexdigest()
34
+ qs = [json.loads(l) for l in raw.decode().splitlines() if l.strip()]
35
+ if a.limit: qs = qs[:a.limit]
36
+ import torch, vllm
37
+ from vllm import LLM, SamplingParams
38
+ env = {"torch": torch.__version__, "cuda_available": torch.cuda.is_available(), "gpu": torch.cuda.get_device_name(0) if torch.cuda.is_available() else None,
39
+ "vllm": vllm.__version__, "model": MODEL, "revision": REV, "questions_sha256": qsha, "n_questions": len(qs), "tries": N_TRIES, "seed": SEED}
40
+ json.dump(env, open(os.path.join(a.out, "env.json"), "w"), indent=1)
41
+ if not torch.cuda.is_available(): sys.exit("CUDA not available")
42
+ llm = LLM(model=MODEL, revision=REV, dtype="bfloat16", max_model_len=2048, gpu_memory_utilization=0.92, seed=SEED, enable_prefix_caching=True)
43
+ t0 = time.time(); recs = open(os.path.join(a.out, "records.jsonl"), "w"); summ = {"fact": [0, 0], "math": [0, 0]}; hist = {"fact": [0]*6, "math": [0]*6}; trunc = 0
44
+ CH = 500
45
+ for i in range(0, len(qs), CH):
46
+ batch = qs[i:i+CH]
47
+ convs = [[{"role": "user", "content": PROMPT[q["kind"]].format(q=q["question"])}] for q in batch]
48
+ sps = [SamplingParams(n=N_TRIES, temperature=0.6, top_p=0.95, max_tokens=MAXTOK[q["kind"]], seed=SEED + i + j) for j, q in enumerate(batch)]
49
+ outs = llm.chat(convs, sps, use_tqdm=False)
50
+ for q, o in zip(batch, outs):
51
+ texts = [c.text for c in o.outputs]; fin = [c.finish_reason for c in o.outputs]
52
+ ok = [bool(fact_ok(t, q["gold"]) if q["kind"] == "fact" else math_ok(t, q["gold"])) for t in texts]
53
+ k = sum(ok); summ[q["kind"]][0] += k; summ[q["kind"]][1] += len(ok); hist[q["kind"]][k] += 1; trunc += sum(f == "length" for f in fin)
54
+ recs.write(json.dumps({"id": q["id"], "kind": q["kind"], "question": q["question"], "gold": q["gold"], "outputs": texts, "correct": ok, "finish": fin, "n_correct": k}) + "\n")
55
+ recs.flush()
56
+ open(os.path.join(a.out, "progress.txt"), "w").write(f"{min(i+CH, len(qs))}/{len(qs)} questions {time.time()-t0:.0f}s\n")
57
+ recs.close()
58
+ s = {"accuracy_per_try": {k: (v[0]/v[1] if v[1] else None) for k, v in summ.items()},
59
+ "questions_by_n_correct_of_5": hist, "truncated_answers": trunc, "seconds": round(time.time()-t0), **env}
60
+ json.dump(s, open(os.path.join(a.out, "summary.json"), "w"), indent=1); print(json.dumps(s))
61
+
62
+ if __name__ == "__main__":
63
+ main()
eval/job_f3.py ADDED
@@ -0,0 +1,207 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Sentinel S1 Job F3 — sign-off of the 32B Sentinel (owner GO 14 Nov 2026 11:47 UK). The 9 checks of rev 4, same methods and pass marks:
2
+ checks 1-4 exactly as Job C2/C4 (IFEval, GSM8K test, NQ-val + SimpleQA, clean endings), check 5 with the adopted stronger search (E3 method:
3
+ 2018 + 2023 Wikipedia, BM25 top 50 each, bge-reranker-v2-m3 top 10, system-named source), and K1-K4 exactly as Job D/D3 on fresh
4
+ conversations. Original model = unmodified Olmo 3.1 32B Instruct; Sentinel = same weights + the Job F2 adapter; identical code, greedy."""
5
+ import argparse, hashlib, json, os, re, shutil, threading, time
6
+ from job_a import SEED, PROMPT, fact_ok, math_ok, norm
7
+ from model32 import MODEL32, REV32
8
+ WIKI_URL, WIKI_MD5 = "https://rgw.cs.uwaterloo.ca/pyserini/indexes/lucene-index.wikipedia-dpr-100w.20210120.d1b9e6.tar.gz", "7b58c08da992b2ea7e96667f0b176651"
9
+ JAR_URL, JAR_SHA256 = "https://repo1.maven.org/maven2/io/anserini/anserini/1.1.1/anserini-1.1.1-fatjar.jar", "69270ba4d160826953347411ce5d7e205ce363766a9cc72ac9da3b945341af83"
10
+ RERANKER, RERANKER_REV = "BAAI/bge-reranker-v2-m3", "953dc6f6f85a1b2dbfca4c34a2796e7dde08d41e"
11
+ PROBES = [("Eiffel Tower Paris", "eiffel"), ("Albert Einstein theory of relativity", "einstein"), ("William Shakespeare playwright Hamlet", "shakespeare"),
12
+ ("Amazon river South America", "amazon"), ("Mount Everest highest mountain", "everest")]
13
+ K_EACH, K_FINAL, MAXTOK, CH, EARLY_TOK, LAST_TOK = 50, 10, 1024, 512, 512, 1024
14
+ NEW_PROMPT = "Answer the question using the Wikipedia passages below. Give a short factual answer only (a few words), then on a new line write 'Source: ' followed by the number of the passage you used.\n\n"
15
+ SEVEN_B_REV4 = {"ifeval": [0.8022, 0.8041], "gsm8k": [0.9143, 0.9128], "unflagged": 0.7506, "flags_wrong": 0.9404, "search_after": 0.2505, "conv_unflagged": 0.8228, "conv_flags_wrong": 0.9247, "conv_chat_flagged": 0.06}
16
+ def flagged(t): return t.lstrip().upper().startswith("UNSURE:")
17
+ def unflag(t):
18
+ s = t.lstrip(); return s[len("UNSURE:"):].strip() if s.upper().startswith("UNSURE:") else t
19
+ def sha_file(p):
20
+ h = hashlib.sha256()
21
+ with open(p, "rb") as f:
22
+ for b in iter(lambda: f.read(1 << 24), b""): h.update(b)
23
+ return h.hexdigest()
24
+ def main():
25
+ ap = argparse.ArgumentParser()
26
+ for k in ("signoff", "convs", "adapter", "w23", "w23meta", "out"): ap.add_argument("--" + k)
27
+ ap.add_argument("--tp", type=int, default=4); ap.add_argument("--limit", type=int, default=0); ap.add_argument("--w18_dir", default=""); ap.add_argument("--jar", default=""); ap.add_argument("--java_home", default=""); ap.add_argument("--rerank_device", default="cuda:0")
28
+ a = ap.parse_args(); t0 = time.time(); prog = lambda m: open(os.path.join(a.out, "progress.txt"), "w").write(f"{m} {time.time()-t0:.0f}s\n")
29
+ disks = {d: shutil.disk_usage(d).free for d in ("/opt/ml", "/tmp") if os.path.exists(d)}; wdir = os.path.join(max(disks, key=disks.get), "scratch_wiki"); os.makedirs(wdir, exist_ok=True)
30
+ dl = {"ok": bool(a.w18_dir), "err": None, "index_dir": a.w18_dir, "jar": a.jar, "java_home": a.java_home}
31
+ def fetch():
32
+ s0 = time.time()
33
+ try:
34
+ import urllib.request, tarfile, jdk
35
+ jar = os.path.join(wdir, "anserini-1.1.1-fatjar.jar"); urllib.request.urlretrieve(JAR_URL, jar)
36
+ if sha_file(jar) != JAR_SHA256: raise ValueError("Anserini jar hash mismatch")
37
+ dl["java_home"] = jdk.install("21", path=os.path.join(wdir, "jdk")); tgz = os.path.join(wdir, "index.tar.gz"); urllib.request.urlretrieve(WIKI_URL, tgz); h = hashlib.md5()
38
+ with open(tgz, "rb") as f:
39
+ for b in iter(lambda: f.read(1 << 24), b""): h.update(b)
40
+ if h.hexdigest() != WIKI_MD5: raise ValueError("index md5 mismatch")
41
+ with tarfile.open(tgz) as t: t.extractall(wdir, filter="data")
42
+ os.remove(tgz); dl["index_dir"] = [os.path.join(wdir, d) for d in os.listdir(wdir) if d.startswith("lucene-index")][0]; dl["jar"] = jar; dl["ok"] = True
43
+ except Exception as e: dl["err"] = repr(e)[:500]
44
+ dl["seconds"] = round(time.time() - s0)
45
+ th = None
46
+ if not a.w18_dir: th = threading.Thread(target=fetch, daemon=True); th.start()
47
+ load = lambda p: [json.loads(l) for l in open(p) if l.strip()]
48
+ sets = {k: load(os.path.join(a.signoff, f"{k}.jsonl")) for k in ("ifeval", "gsm8k_test", "nq_val", "simpleqa")}
49
+ convs = load(os.path.join(a.convs, "conversations.jsonl"))
50
+ if a.limit: sets = {k: v[:a.limit] for k, v in sets.items()}; convs = convs[:a.limit] + [c for c in convs if c["group"].startswith("D_")][:a.limit]
51
+ import torch, vllm
52
+ from vllm import LLM, SamplingParams
53
+ from vllm.lora.request import LoRARequest
54
+ env = {"model": MODEL32, "revision": REV32, "vllm": vllm.__version__, "torch": torch.__version__, "gpus": [torch.cuda.get_device_name(i) for i in range(torch.cuda.device_count())],
55
+ "signoff_sha256": {k: sha_file(os.path.join(a.signoff, f"{k}.jsonl")) for k in sets}, "conversations_sha256": sha_file(os.path.join(a.convs, "conversations.jsonl")),
56
+ "adapter_sha256": {f: sha_file(os.path.join(a.adapter, f)) for f in sorted(os.listdir(a.adapter)) if os.path.isfile(os.path.join(a.adapter, f))}, "tensor_parallel": a.tp}
57
+ json.dump(env, open(os.path.join(a.out, "env.json"), "w"), indent=1)
58
+ llm = LLM(model=MODEL32, revision=REV32, dtype="bfloat16", tensor_parallel_size=a.tp, max_model_len=8192, gpu_memory_utilization=0.85, seed=SEED, enforce_eager=True, # D72: A10G memory fix (no CUDA graphs)
59
+ enable_lora=True, max_lora_rank=16, max_loras=1, enable_prefix_caching=True)
60
+ lora = LoRARequest("sentinel", 1, a.adapter)
61
+ # ===== checks 1-4: exactly as Job C2/C4
62
+ sp = SamplingParams(temperature=0.0, max_tokens=MAXTOK, seed=SEED)
63
+ from ifeval.utils import process_results
64
+ recs = {}
65
+ for arm in ("base", "sentinel"):
66
+ out = []
67
+ jobs = [("ifeval", r, r["prompt"]) for r in sets["ifeval"]] + [(k, r, PROMPT[r["kind"]].format(q=r["question"])) for k in ("gsm8k_test", "nq_val", "simpleqa") for r in sets[k]]
68
+ for i in range(0, len(jobs), CH):
69
+ part = jobs[i:i+CH]
70
+ res = llm.chat([[{"role": "user", "content": p}] for _, _, p in part], sp, lora_request=lora if arm == "sentinel" else None, use_tqdm=False)
71
+ for (k, r, _), o in zip(part, res):
72
+ t = o.outputs[0].text; fin = o.outputs[0].finish_reason
73
+ rec = {"set": k, "id": r.get("id", r.get("key")), "reply": t, "finish": fin, "flagged": flagged(t)}
74
+ if k == "ifeval": rec["correct"] = bool(process_results(r, [t])["prompt_level_strict_acc"])
75
+ else: rec["correct"] = bool(fact_ok(unflag(t), r["gold"]) if r["kind"] == "fact" else math_ok(unflag(t), r["gold"]))
76
+ out.append(rec)
77
+ prog(f"signoff {arm} {min(i+CH, len(jobs))}/{len(jobs)}")
78
+ recs[arm] = out
79
+ with open(os.path.join(a.out, f"signoff_{arm}.jsonl"), "w") as f:
80
+ for x in out: f.write(json.dumps(x) + "\n")
81
+ def acc(rows): return sum(r["correct"] for r in rows) / len(rows) if rows else None
82
+ B, S = recs["base"], recs["sentinel"]; sel = lambda rows, ks: [r for r in rows if r["set"] in ks]
83
+ QA = ("gsm8k_test", "nq_val", "simpleqa"); FACT = ("nq_val", "simpleqa")
84
+ c1 = {"base": acc(sel(B, ("ifeval",))), "sentinel": acc(sel(S, ("ifeval",)))}; c1["pass"] = c1["sentinel"] >= c1["base"] - 0.02
85
+ c2 = {"base": acc(sel(B, ("gsm8k_test",))), "sentinel": acc(sel(S, ("gsm8k_test",)))}; c2["pass"] = c2["sentinel"] >= c2["base"] - 0.02
86
+ def c3for(ks):
87
+ b = sel(B, ks); s = sel(S, ks); un = [r for r in s if not r["flagged"]]; wrong = [r for r in s if not r["correct"]]
88
+ return {"base_accuracy": acc(b), "sentinel_unflagged_accuracy": acc(un), "sentinel_unflagged_n": len(un), "sentinel_flag_rate": sum(r["flagged"] for r in s) / len(s),
89
+ "sentinel_flag_rate_on_its_wrong_answers": (sum(r["flagged"] for r in wrong) / len(wrong)) if wrong else None, "sentinel_overall_accuracy": acc(s)}
90
+ c3 = c3for(QA); c3["pass"] = (c3["sentinel_unflagged_accuracy"] is not None and c3["sentinel_unflagged_accuracy"] >= c3["base_accuracy"] + 0.10
91
+ and c3["sentinel_flag_rate_on_its_wrong_answers"] is not None and c3["sentinel_flag_rate_on_its_wrong_answers"] >= 0.50)
92
+ c3_breakdown = {"facts_only": c3for(FACT), "maths_only": c3for(("gsm8k_test",)), "nq_val": c3for(("nq_val",)), "simpleqa": c3for(("simpleqa",))}
93
+ c4 = {"sentinel_natural_end_rate": sum(r["finish"] == "stop" for r in S) / len(S), "base_natural_end_rate": sum(r["finish"] == "stop" for r in B) / len(B)}; c4["pass"] = c4["sentinel_natural_end_rate"] >= 0.95
94
+ summary = {"check1_instructions_ifeval": c1, "check2_maths_gsm8k": c2, "check3_knows_when_unsure": c3, "check3_breakdown_D18": c3_breakdown, "check4_clean_replies": c4}
95
+ json.dump(summary, open(os.path.join(a.out, "summary.json"), "w"), indent=1)
96
+ # ===== K1-K4: exactly as Job D/D3
97
+ crec = {}
98
+ for arm in ("base", "sentinel"):
99
+ msgs = {c["id"]: [] for c in convs}; turns = {c["id"]: [] for c in convs}
100
+ for step in range(max(len(c["user_turns"]) for c in convs)):
101
+ act = [c for c in convs if len(c["user_turns"]) > step]
102
+ for c in act: msgs[c["id"]].append({"role": "user", "content": c["user_turns"][step]})
103
+ sps = [SamplingParams(temperature=0.0, seed=SEED, max_tokens=LAST_TOK if step == len(c["user_turns"]) - 1 else EARLY_TOK) for c in act]
104
+ res = llm.chat([msgs[c["id"]] for c in act], sps, lora_request=lora if arm == "sentinel" else None, use_tqdm=False)
105
+ for c, o in zip(act, res):
106
+ t = o.outputs[0].text; msgs[c["id"]].append({"role": "assistant", "content": t}); turns[c["id"]].append({"reply": t, "finish": o.outputs[0].finish_reason})
107
+ prog(f"conversations {arm} turn {step+1}")
108
+ out = []
109
+ for c in convs:
110
+ last = turns[c["id"]][-1]; t = last["reply"]
111
+ r = {"id": c["id"], "group": c["group"], "kind": c["kind"], "reply": t, "finish": last["finish"], "flagged": flagged(t), "unsure_anywhere": "UNSURE:" in t.upper(),
112
+ "earlier_turns_flagged": sum(flagged(x["reply"]) for x in turns[c["id"]][:-1])}
113
+ if c["kind"]: r["correct"] = bool(fact_ok(unflag(t), c["gold"]) if c["kind"] == "fact" else math_ok(unflag(t), c["gold"]))
114
+ out.append(r)
115
+ crec[arm] = out
116
+ with open(os.path.join(a.out, f"conv_records_{arm}.jsonl"), "w") as f:
117
+ for x in out: f.write(json.dumps(x) + "\n")
118
+ with open(os.path.join(a.out, f"conv_transcripts_{arm}.jsonl"), "w") as f:
119
+ for c in convs: f.write(json.dumps({"id": c["id"], "messages": msgs[c["id"]]}) + "\n")
120
+ CB, CS = crec["base"], crec["sentinel"]; q = lambda rows: [r for r in rows if r["kind"]]
121
+ un = [r for r in q(CS) if not r["flagged"]]; wr = [r for r in q(CS) if not r["correct"]]; D = [r for r in CS if r["group"].startswith("D_")]
122
+ K1 = {"sentinel_unflagged_accuracy": acc(un), "base_accuracy": acc(q(CB))}; K1["pass"] = K1["sentinel_unflagged_accuracy"] is not None and K1["sentinel_unflagged_accuracy"] >= K1["base_accuracy"] + 0.10
123
+ K2 = {"sentinel_flag_rate_on_its_wrong": (sum(r["flagged"] for r in wr) / len(wr)) if wr else None}; K2["pass"] = K2["sentinel_flag_rate_on_its_wrong"] is not None and K2["sentinel_flag_rate_on_its_wrong"] >= 0.50
124
+ K3 = {"sentinel_clean_end_rate": sum(r["finish"] == "stop" for r in CS) / len(CS), "base_clean_end_rate": sum(r["finish"] == "stop" for r in CB) / len(CB)}; K3["pass"] = K3["sentinel_clean_end_rate"] >= 0.95
125
+ K4 = {"sentinel_flag_rate_no_fact_chats": sum(r["flagged"] for r in D) / len(D), "n": len(D)}; K4["pass"] = K4["sentinel_flag_rate_no_fact_chats"] <= 0.10
126
+ summary.update({"K1_unflagged_beats_original_by_10pts": K1, "K2_flags_half_its_wrong": K2, "K3_clean_endings": K3, "K4_no_false_flags_on_chat": K4,
127
+ "conv_reported": {"sentinel_overall_accuracy": acc(q(CS)), "sentinel_flag_rate": sum(r["flagged"] for r in q(CS)) / len(q(CS)), "earlier_turns_flagged": sum(r["earlier_turns_flagged"] for r in CS),
128
+ "maths_flag_rate_on_wrong": (lambda w: sum(r["flagged"] for r in w) / len(w) if w else None)([r for r in q(CS) if r["kind"] == "math" and not r["correct"]])}})
129
+ json.dump(summary, open(os.path.join(a.out, "summary.json"), "w"), indent=1)
130
+ # ===== check 5: stronger search (E3 method) on the fact questions Sentinel flags
131
+ prog("waiting for 2018 index")
132
+ if th: th.join()
133
+ c5 = {"index_2018": {k: v for k, v in dl.items() if k != "java_home"}}
134
+ man = json.load(open(os.path.join(a.w23meta, "index_manifest.json"))); bad = [k for k, v in man.items() if sha_file(os.path.join(a.w23, k)) != v["sha256"]]
135
+ c5["w23_index_files_verified"] = f"{len(man) - len(bad)}/{len(man)}"
136
+ ffs = [r for r in sel(S, FACT) if r["flagged"]]; ffc = [r for r in q(CS) if r["kind"] == "fact" and r["flagged"]]; c5["flagged_fact_questions"] = len(ffs); c5["flagged_fact_conversations"] = len(ffc)
137
+ if dl["ok"] and not bad and ffs:
138
+ os.environ["JAVA_HOME"] = dl["java_home"]; os.environ["PATH"] = os.path.join(dl["java_home"], "bin") + os.pathsep + os.environ["PATH"]
139
+ import jnius_config; jnius_config.add_options("-Xmx8g"); jnius_config.set_classpath(dl["jar"])
140
+ from jnius import autoclass
141
+ JS = autoclass("io.anserini.search.SimpleSearcher"); AL = autoclass("java.util.ArrayList"); SR = {"w18": JS(dl["index_dir"]), "w23": JS(a.w23)}
142
+ def passage(s, docid):
143
+ raw = SR[s].doc(docid).get("raw") or ""
144
+ try: c = json.loads(raw).get("contents", raw)
145
+ except Exception: c = raw
146
+ first, _, rest = c.partition("\n"); return {"title": first.strip().strip('"'), "text": rest.strip(), "src": s}
147
+ def search(s, texts, k):
148
+ out = []
149
+ for i in range(0, len(texts), 512):
150
+ part = texts[i:i+512]; ql, ids = AL(), AL()
151
+ for j, t in enumerate(part): ql.add(t); ids.add(str(j))
152
+ res = SR[s].batch_search(ql, ids, int(k), 8); out += [[passage(s, h.docid) for h in res.get(str(j))] for j in range(len(part))]
153
+ return out
154
+ c5["probe_hits"] = {s: sum(any(w in d["title"].lower() for d in row) for (p, w), row in zip(PROBES, search(s, [p for p, _ in PROBES], 3))) for s in SR}
155
+ if min(c5["probe_hits"].values()) >= 4:
156
+ qtext = {r["id"]: r for k in FACT for r in sets[k]}; cmap = {c["id"]: c for c in convs}
157
+ items = [("signoff", r, qtext[r["id"]]["question"], qtext[r["id"]]["gold"]) for r in ffs] + [("conversation", r, cmap[r["id"]]["user_turns"][-1], cmap[r["id"]]["gold"]) for r in ffc]
158
+ texts = [it[2] for it in items]; prog("searching")
159
+ c18 = search("w18", texts, K_EACH); c23 = search("w23", texts, K_EACH); cands = []
160
+ for x, y in zip(c18, c23):
161
+ seen, row = set(), []
162
+ for d in x + y:
163
+ key = (d["title"], d["text"][:200])
164
+ if key not in seen: seen.add(key); row.append(d)
165
+ cands.append(row)
166
+ prog("reranking")
167
+ from transformers import AutoTokenizer, AutoModelForSequenceClassification
168
+ tok = AutoTokenizer.from_pretrained(RERANKER, revision=RERANKER_REV); rr = AutoModelForSequenceClassification.from_pretrained(RERANKER, revision=RERANKER_REV, dtype=torch.float16 if a.rerank_device.startswith("cuda") else torch.float32).to(a.rerank_device).eval()
169
+ top = []
170
+ for qi, (qq, row) in enumerate(zip(texts, cands)):
171
+ pairs = [[qq, f"{d['title']}: {d['text']}"] for d in row]; scores = []
172
+ for i in range(0, len(pairs), 32):
173
+ enc = tok(pairs[i:i+32], padding=True, truncation=True, max_length=384, return_tensors="pt").to(a.rerank_device)
174
+ with torch.no_grad(): scores += rr(**enc).logits.view(-1).float().tolist()
175
+ top.append([row[i] for i in sorted(range(len(row)), key=lambda i: -scores[i])[:K_FINAL]])
176
+ if qi % 200 == 0: prog(f"reranking {qi}/{len(texts)}")
177
+ del rr
178
+ prog("answering with search")
179
+ prompts = [NEW_PROMPT + "\n\n".join(f"[{j+1}] {d['title']}: {d['text'][:600]}" for j, d in enumerate(row)) + f"\n\nQuestion: {qq}" for qq, row in zip(texts, top)]
180
+ res = llm.chat([[{"role": "user", "content": p}] for p in prompts], SamplingParams(temperature=0.0, max_tokens=128, seed=SEED), lora_request=lora, use_tqdm=False)
181
+ def cite(ans, row):
182
+ na = norm(ans)
183
+ if not na: return None
184
+ for j, d in enumerate(row):
185
+ if f" {na} " in f" {norm(d['title'] + ' ' + d['text'][:600])} ": return j + 1
186
+ return None
187
+ srecs = []
188
+ for (kind, r, qq, gold), row, o in zip(items, top, res):
189
+ t = o.outputs[0].text; ans = unflag(t.split("Source:")[0].strip()); n = cite(ans, row)
190
+ srecs.append({"kind": kind, "id": r["id"], "guess_correct": r["correct"], "reply": t, "answer": ans, "search_correct": bool(fact_ok(ans, gold)),
191
+ "system_source": row[n - 1]["title"] if n else "no source found", "passages": [[d["src"], d["title"]] for d in row]})
192
+ with open(os.path.join(a.out, "search_records.jsonl"), "w") as f:
193
+ for x in srecs: f.write(json.dumps(x) + "\n")
194
+ ss = [x for x in srecs if x["kind"] == "signoff"]; sc = [x for x in srecs if x["kind"] == "conversation"]
195
+ c5["guess_accuracy_on_flagged"] = sum(x["guess_correct"] for x in ss) / len(ss); c5["search_accuracy_on_flagged"] = sum(x["search_correct"] for x in ss) / len(ss)
196
+ c5["answers_with_a_source"] = sum(x["system_source"] != "no source found" for x in ss) / len(ss)
197
+ c5["conversation_guess"] = (sum(x["guess_correct"] for x in sc) / len(sc)) if sc else None; c5["conversation_search"] = (sum(x["search_correct"] for x in sc) / len(sc)) if sc else None
198
+ c5["pass"] = c5["search_accuracy_on_flagged"] > c5["guess_accuracy_on_flagged"]
199
+ else: c5["pass"] = None; c5["note"] = "index probe gate failed (<4/5): search not evaluated"
200
+ else: c5["pass"] = None; c5["note"] = "2018 index download failed, 2023 index files did not verify, or nothing flagged"
201
+ summary["check5_search_when_unsure"] = c5
202
+ keys = ("check1_instructions_ifeval", "check2_maths_gsm8k", "check3_knows_when_unsure", "check4_clean_replies", "check5_search_when_unsure",
203
+ "K1_unflagged_beats_original_by_10pts", "K2_flags_half_its_wrong", "K3_clean_endings", "K4_no_false_flags_on_chat")
204
+ summary["checks_passed"] = f"{sum(summary[k].get('pass') is True for k in keys)}/9"; summary["all_pass"] = all(summary[k].get("pass") is True for k in keys)
205
+ summary["for_comparison_7b_rev4"] = SEVEN_B_REV4; summary["seconds"] = round(time.time() - t0); summary["env"] = env
206
+ json.dump(summary, open(os.path.join(a.out, "summary.json"), "w"), indent=1); prog("done"); print(json.dumps({"checks_passed": summary["checks_passed"]}))
207
+ if __name__ == "__main__": main()
eval/model32.py ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ """Job F base model (owner GO 14 Nov 2026 11:47 UK): Olmo 3.1 32B Instruct, Ai2, Apache 2.0, pinned revision."""
2
+ MODEL32, REV32 = "allenai/Olmo-3.1-32B-Instruct", "ac0587e4a7744a551c059d8cd17ba220bc940dae"
eval/requirements.txt ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ vllm==0.31.0
2
+ langdetect==1.0.9
3
+ immutabledict==4.3.1
4
+ nltk==3.10.4
5
+ pyjnius==1.7.0
6
+ install-jdk==1.1.0
results/evidence_run_summary.json ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "accuracy_per_try": {
3
+ "fact": 0.32488,
4
+ "math": 0.95676
5
+ },
6
+ "questions_by_n_correct_of_5": {
7
+ "fact": [
8
+ 3084,
9
+ 167,
10
+ 138,
11
+ 113,
12
+ 150,
13
+ 1348
14
+ ],
15
+ "math": [
16
+ 115,
17
+ 40,
18
+ 39,
19
+ 52,
20
+ 125,
21
+ 4629
22
+ ]
23
+ },
24
+ "truncated_answers": 440,
25
+ "casual_all_turns_clean": 1381,
26
+ "seconds": 6374,
27
+ "model": "allenai/Olmo-3.1-32B-Instruct",
28
+ "revision": "ac0587e4a7744a551c059d8cd17ba220bc940dae",
29
+ "vllm": "0.31.0",
30
+ "torch": "2.13.0+cu130",
31
+ "gpus": [
32
+ "NVIDIA A10G",
33
+ "NVIDIA A10G",
34
+ "NVIDIA A10G",
35
+ "NVIDIA A10G"
36
+ ],
37
+ "questions_sha256": "f786f4b054326e9febab6fb275bf7ad18ce8cdffb829f57126fac51734c89996",
38
+ "casual_sha256": "b43705b1a7187f3d6c7dff73574dc876abcd30daf13d395b2d47e83b1fb59672",
39
+ "n_questions": 10000,
40
+ "n_casual": 1445,
41
+ "tries": 5,
42
+ "seed": 20261112,
43
+ "tensor_parallel": 4
44
+ }
results/signoff_summary.json ADDED
@@ -0,0 +1,151 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "check1_instructions_ifeval": {
3
+ "base": 0.8558225508317929,
4
+ "sentinel": 0.8502772643253235,
5
+ "pass": true
6
+ },
7
+ "check2_maths_gsm8k": {
8
+ "base": 0.9469294920394238,
9
+ "sentinel": 0.9469294920394238,
10
+ "pass": true
11
+ },
12
+ "check3_knows_when_unsure": {
13
+ "base_accuracy": 0.2825499729875743,
14
+ "sentinel_unflagged_accuracy": 0.681782945736434,
15
+ "sentinel_unflagged_n": 2580,
16
+ "sentinel_flag_rate": 0.7212317666126418,
17
+ "sentinel_flag_rate_on_its_wrong_answers": 0.8752469229600365,
18
+ "sentinel_overall_accuracy": 0.28892490545651,
19
+ "pass": true
20
+ },
21
+ "check3_breakdown_D18": {
22
+ "facts_only": {
23
+ "base_accuracy": 0.17212701612903225,
24
+ "sentinel_unflagged_accuracy": 0.48588709677419356,
25
+ "sentinel_unflagged_n": 1488,
26
+ "sentinel_flag_rate": 0.8125,
27
+ "sentinel_flag_rate_on_its_wrong_answers": 0.8825065274151436,
28
+ "sentinel_overall_accuracy": 0.17956149193548387
29
+ },
30
+ "maths_only": {
31
+ "base_accuracy": 0.9469294920394238,
32
+ "sentinel_unflagged_accuracy": 0.9487179487179487,
33
+ "sentinel_unflagged_n": 1092,
34
+ "sentinel_flag_rate": 0.17210007581501138,
35
+ "sentinel_flag_rate_on_its_wrong_answers": 0.2,
36
+ "sentinel_overall_accuracy": 0.9469294920394238
37
+ },
38
+ "nq_val": {
39
+ "base_accuracy": 0.3290858725761773,
40
+ "sentinel_unflagged_accuracy": 0.5903508771929824,
41
+ "sentinel_unflagged_n": 1140,
42
+ "sentinel_flag_rate": 0.6842105263157895,
43
+ "sentinel_flag_rate_on_its_wrong_answers": 0.8053355564818675,
44
+ "sentinel_overall_accuracy": 0.33545706371191136
45
+ },
46
+ "simpleqa": {
47
+ "base_accuracy": 0.041146555709662504,
48
+ "sentinel_unflagged_accuracy": 0.14367816091954022,
49
+ "sentinel_unflagged_n": 348,
50
+ "sentinel_flag_rate": 0.9195561719833565,
51
+ "sentinel_flag_rate_on_its_wrong_answers": 0.9275291828793775,
52
+ "sentinel_overall_accuracy": 0.04946833102172908
53
+ }
54
+ },
55
+ "check4_clean_replies": {
56
+ "sentinel_natural_end_rate": 0.9961208656594528,
57
+ "base_natural_end_rate": 0.995712535728869,
58
+ "pass": true
59
+ },
60
+ "K1_unflagged_beats_original_by_10pts": {
61
+ "sentinel_unflagged_accuracy": 0.7785714285714286,
62
+ "base_accuracy": 0.55,
63
+ "pass": true
64
+ },
65
+ "K2_flags_half_its_wrong": {
66
+ "sentinel_flag_rate_on_its_wrong": 0.7876712328767124,
67
+ "pass": true
68
+ },
69
+ "K3_clean_endings": {
70
+ "sentinel_clean_end_rate": 0.9971428571428571,
71
+ "base_clean_end_rate": 0.9885714285714285,
72
+ "pass": true
73
+ },
74
+ "K4_no_false_flags_on_chat": {
75
+ "sentinel_flag_rate_no_fact_chats": 0.06,
76
+ "n": 50,
77
+ "pass": true
78
+ },
79
+ "conv_reported": {
80
+ "sentinel_overall_accuracy": 0.5133333333333333,
81
+ "sentinel_flag_rate": 0.5333333333333333,
82
+ "earlier_turns_flagged": 0,
83
+ "maths_flag_rate_on_wrong": 0.3333333333333333
84
+ },
85
+ "check5_search_when_unsure": {
86
+ "index_2018": {
87
+ "ok": true,
88
+ "err": null,
89
+ "index_dir": "/opt/ml/scratch_wiki/lucene-index.wikipedia-dpr-100w.20210120.d1b9e6",
90
+ "jar": "/opt/ml/scratch_wiki/anserini-1.1.1-fatjar.jar",
91
+ "seconds": 251
92
+ },
93
+ "w23_index_files_verified": "80/80",
94
+ "flagged_fact_questions": 6448,
95
+ "flagged_fact_conversations": 149,
96
+ "probe_hits": {
97
+ "w18": 4,
98
+ "w23": 4
99
+ },
100
+ "guess_accuracy_on_flagged": 0.10887096774193548,
101
+ "search_accuracy_on_flagged": 0.4522332506203474,
102
+ "answers_with_a_source": 0.7689205955334988,
103
+ "conversation_guess": 0.2348993288590604,
104
+ "conversation_search": 0.31543624161073824,
105
+ "pass": true
106
+ },
107
+ "checks_passed": "9/9",
108
+ "all_pass": true,
109
+ "for_comparison_7b_rev4": {
110
+ "ifeval": [
111
+ 0.8022,
112
+ 0.8041
113
+ ],
114
+ "gsm8k": [
115
+ 0.9143,
116
+ 0.9128
117
+ ],
118
+ "unflagged": 0.7506,
119
+ "flags_wrong": 0.9404,
120
+ "search_after": 0.2505,
121
+ "conv_unflagged": 0.8228,
122
+ "conv_flags_wrong": 0.9247,
123
+ "conv_chat_flagged": 0.06
124
+ },
125
+ "seconds": 17711,
126
+ "env": {
127
+ "model": "allenai/Olmo-3.1-32B-Instruct",
128
+ "revision": "ac0587e4a7744a551c059d8cd17ba220bc940dae",
129
+ "vllm": "0.31.0",
130
+ "torch": "2.13.0+cu130",
131
+ "gpus": [
132
+ "NVIDIA A10G",
133
+ "NVIDIA A10G",
134
+ "NVIDIA A10G",
135
+ "NVIDIA A10G"
136
+ ],
137
+ "signoff_sha256": {
138
+ "ifeval": "2b323e8baa968ec6581eb82460abfd274c37d7d986bc2b6b66401676a1d8cc23",
139
+ "gsm8k_test": "8f656bc76a06108bedb7625c7313b1cd57d9ff2110f213b3fd068f509e84944b",
140
+ "nq_val": "d298b705fc022ff838ad48ea61bd0ec79320b978e471df5b02b89be2834b2161",
141
+ "simpleqa": "efc7670e0bf22ebb55c3f8a944f69a2350bf9a861341233a5ef67106fa1195b4"
142
+ },
143
+ "conversations_sha256": "f8c823e9dda5e1c8e1120f0be0ad6cb3f2e8f819e13cd077bcda36e33a80d1c3",
144
+ "adapter_sha256": {
145
+ "README.md": "d8f69bfae9dbc754d70dfdd1fd58018af59d43e8893079a776afad04a57a909c",
146
+ "adapter_config.json": "f7a551b72ca4fdfea13b982eded92a81102e2473dd24707277e61ded456c09de",
147
+ "adapter_model.safetensors": "546a6494b46670c2bed2a2a30bd15a87d2ed691da85c57dddc6d0bee2ccdce11"
148
+ },
149
+ "tensor_parallel": 4
150
+ }
151
+ }