timaeus/lang5_probes

Selected Probes

Each probe is a CSV with prompt, prompt_len, and target columns. All targets are 0/1 integers unless noted. All datasets are balanced (50/50) unless noted.

---

5 — hist_fig_ismale

Entries: 5,000 | Avg prompt length: 20 chars | Max: 70 chars

Prompts: Historical figure names (e.g. "Margaret of Clisson", "Billy Mays").

Target: 1 = male, 0 = female — 50% / 50%

---

6 — hist_fig_isamerican

Entries: 5,000 | Avg prompt length: 17 chars | Max: 65 chars

Prompts: Historical figure names (same pool as above).

Target: 1 = American, 0 = non-American — 50% / 50%

---

22 — headline_isobama

Entries: 1,656 | Avg prompt length: 57 chars | Max: 108 chars

Prompts: News headlines (e.g. "Obama Sticks to a Deadline in Iraq .").

Target: 1 = headline is about Obama, 0 = not about Obama — 50% / 50%

Note: target column was originally True/False; converted to 1/0.

---

36 — sciq_tf

Entries: 5,000 | Avg prompt length: 94 chars | Max: 427 chars

Prompts: Science question + answer pairs formatted as Q: ...\nA: ....

Target: 1 = answer is correct, 0 = answer is incorrect — 50% / 50%

---

49 — cm_isshort

Entries: 5,000 | Avg prompt length: 981 chars | Max: 9,851 chars

Prompts: Reddit-style posts and stories (long-form text, variable topics).

Target: 1 = short text, 0 = long text — 50% / 50%

Note: target column was originally True/False; converted to 1/0.

---

51 — just_is

Entries: 5,000 | Avg prompt length: 103 chars | Max: 327 chars

Prompts: Short descriptions of a person's action or decision, written in first or third person.

Target: 1 = action is justified, 0 = not justified — 50% / 50%

---

65 — high-school

Entries: 3,200 | Avg prompt length: 108 chars | Max: 786 chars

Prompts: Miscellaneous text excerpts, some of which contain a "high school" bigram near the end.

Target: 1 = ends with "high school" bigram, 0 = does not — 50% / 50%

---

92 — glue_sst2

Entries: 5,000 | Avg prompt length: 53 chars | Max: 256 chars

Prompts: Movie review excerpts (from the SST-2 benchmark).

Target: 1 = positive sentiment, 0 = negative sentiment — 50% / 50%

---

96 — spam_is

Entries: 1,494 | Avg prompt length: 108 chars | Max: 791 chars

Prompts: SMS/text messages of varying content and length.

Target: 1 = spam, 0 = not spam — 50% / 50%

---

122 — us_timezone_Los_Angeles

Entries: 3,332 | Avg prompt length: 13 chars | Max: 89 chars

Prompts: US location names (cities, towns, neighborhoods).

Target: 1 = Pacific timezone (Los Angeles), 0 = other timezone — 50% / 50%

---

129 — arith_mc_A

Entries: 1,494 | Avg prompt length: 53 chars | Max: 67 chars

Prompts: Arithmetic questions with four multiple-choice answers, formatted as:

Calculate 63 - 5 =
A.57 B.59 C.68 D.58
Answer: A

"Answer: A" appended to each prompt to steer the model toward option A.

Target: 1 = correct answer is A, 0 = correct answer is not A — 50% / 50%

---

139 — news_class_Politics

Entries: 2,500 | Avg prompt length: 237 chars | Max: 745 chars

Prompts: News article headlines + body snippets from various sections.

Target: 1 = politics section, 0 = not politics — 50% / 50%

---

145 — disease_class_digestive system diseases

Entries: 1,052 | Avg prompt length: 1,236 chars | Max: 2,939 chars

Prompts: Abstracts from medical/science articles about diseases.

Target: 1 = digestive system disease, 0 = other disease category — 50% / 50%

---

150 — twt_emotion_sadness

Entries: 2,500 | Avg prompt length: 76 chars | Max: 152 chars

Prompts: Tweets of varying emotional tone.

Target: 1 = tweet expresses sadness, 0 = other emotion — 50% / 50%

---

159 — code_Python

Entries: 3,672 | Avg prompt length: ~3,500 chars (median) | Max: 24,999 chars

Prompts: Code snippets of varying length and language. Filtered to <25k chars (dropped 328 rows that were large auto-generated files, C headers, etc.); original max was 853k chars and outliers skewed slightly non-Python.

Target: 1 = Python code, 0 = other language — 50.5% / 49.5%