TeamNull commited on
Commit
e47dec8
·
verified ·
1 Parent(s): 745ec66

Restructure card: quick start and scores first

Browse files
Files changed (1) hide show
  1. README.md +170 -215
README.md CHANGED
@@ -20,76 +20,74 @@ configs:
20
  path: data/deeparithmark-1.jsonl
21
  ---
22
 
 
23
 
24
- ![image](https://cdn-uploads.huggingface.co/production/uploads/69c842686cf758859915159c/q3tmOzuyvXpjF6wOY3OrV.png)
25
-
26
- **The long-horizon tier of [ArithMark 3.0](https://huggingface.co/datasets/AxiomicLabs/Arithmark-3.0).**
27
-
28
- > **Unofficial.** Deep Arithmark is an independent, community-built companion
29
- > tier. It is not affiliated with, endorsed by, or published by AxiomicLabs, the
30
- > authors of ArithMark 3.0, and it is not an official ArithMark release. It
31
- > follows their task contract and credits them throughout; any errors are the
32
- > author's. If you use this benchmark, cite ArithMark 3.0 as well.
33
-
34
- ArithMark 3.0 asks a base model to finish a one- or two-step word problem whose
35
- four continuations are **adjacent integers** (43 / 44 / 45 / 46). To reject a
36
- wrong option you only need to know the answer is "about 44", so a magnitude
37
- prior is enough and arithmetic is optional.
38
-
39
- Deep Arithmark keeps ArithMark's contract exactly — one unfinished sentence, four
40
- continuations with a leading space, one correct index, scored by
41
- length-normalised continuation log-likelihood, chance 25 % — and changes three
42
- things:
43
-
44
- 1. **Depth.** Every item needs **2–5 dependent operations** (250 items per tier).
45
- Later operations consume earlier results, so no single memorised template
46
- yields the continuation.
47
- 2. **Decoys with provenance.** No decoy is "answer ± 1". Every decoy is the exact
48
- value some specific plausible wrong reasoning produces — the state after step
49
- 2 instead of step 5, the percentage on the wrong base, the divisor `n` instead
50
- of `n + 1`, the unit conversion left out. `metadata.option_sources` records
51
- which mistake each ending encodes, so error analysis can ask *which* step a
52
- model botches.
53
- 3. **Balanced on three axes.** The correct slot, the answer's *numeric rank* and
54
- the *roundness class* of all four options are all balanced, so "pick the
55
- largest" and "pick the roundest number" are worth exactly 25 %.
56
-
57
- Full design notes, the fairness audit and the measurement discussion are in
58
- `DEEPARITHMARK_DESIGN.md`, which ships in this repository next to the generator
59
- (`deeparithmark.py`) and the scorer (`bench_deeparithmark.py`).
60
-
61
- ## At a glance
62
-
63
- | Property | Value |
64
  | --- | ---: |
65
  | Examples | 1,000 |
66
- | Answer choices per example | 4 |
67
- | Random-choice accuracy | 25 % |
68
- | Templates | 15 |
69
- | Topic labels | 14 |
70
  | Depth tiers | 2, 3, 4, 5 — 250 each |
 
71
  | Mean prompt length | 29.2 words (max 43) |
72
- | Mean ending length | 9.7 characters |
73
- | Language | English |
74
- | Data source | Synthetic (generated, deterministic seed) |
75
- | Dataset file | `data/deeparithmark-1.jsonl` |
76
- | SHA-256 | `8a54eb90b66b142b7fea422b34d851746d1cabc476dcba906c7f3ad5ddc66ee0` |
77
 
78
- ## Repository contents
79
 
80
- | File | What it is |
81
- | --- | --- |
82
- | `data/deeparithmark-1.jsonl` | The dataset — 1,000 items, one JSON object per line |
83
- | `README.md` | This card |
84
- | `deeparithmark.py` | Deterministic generator, plus `selftest` and the shortcut `audit` |
85
- | `bench_deeparithmark.py` | Scorer with per-depth and per-template breakdowns and an ArithMark 3.0 comparison |
86
- | `DEEPARITHMARK_DESIGN.md` | Design rationale, fairness audit, measurement discussion |
 
 
 
 
 
 
 
 
 
 
87
 
88
- ## Task format
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
89
 
90
- Identical to ArithMark 3.0, so `bencharithmark-3.py --data-path
91
- data/deeparithmark-1.jsonl` scores it unchanged (verified against that script's
92
- own loader).
93
 
94
  ```json
95
  {
@@ -119,80 +117,17 @@ own loader).
119
  }
120
  ```
121
 
122
- The correct continuation is ` 26 crates.` — `420 + 260 − 210 − 40 = 430`, and
123
- `430 // 16 = 26` full crates. Every wrong ending is the result of a named
124
- mistake, not a perturbation of 26.
125
-
126
- ## Composition
127
-
128
- ### Depth
129
-
130
- | Depth | Examples | Share | Typical item |
131
- | ---: | ---: | ---: | --- |
132
- | 2 | 250 | 25.0 % | percent off a price, then a fee |
133
- | 3 | 250 | 25.0 % | unit conversion, add, subtract |
134
- | 4 | 250 | 25.0 % | receive / ship / damage / pack |
135
- | 5 | 250 | 25.0 % | two percentage days, then palletise |
136
-
137
- `depth` equals the number of operations in `metadata.solution_steps` for every
138
- item; the build fails if that is not true.
139
-
140
- ### Templates
141
-
142
- | Template | Depth | Examples | Topic |
143
- | --- | ---: | ---: | --- |
144
- | `money_discount_fee` | 2 | 122 | money_percent |
145
- | `time_clock_end` | 2 | 128 | time |
146
- | `measure_cm_cut` | 3 | 33 | measurement |
147
- | `share_then_split` | 3 | 44 | fractions |
148
- | `two_rate_pay` | 3 | 45 | rate_pay |
149
- | `stock_percent_pack` | 3 | 48 | percent |
150
- | `trip_minutes` | 3 | 33 | rate_time |
151
- | `schedule_sessions` | 3 | 47 | time |
152
- | `inventory_boxes` | 4 | 62 | multi_step_inventory |
153
- | `mix_split` | 4 | 63 | money |
154
- | `tank_fill` | 4 | 62 | rate |
155
- | `discount_tax_coupon_split` | 4 | 63 | percent_money |
156
- | `salary_overtime_tax` | 5 | 84 | money_rate |
157
- | `inventory_percent_chain` | 5 | 83 | percent_multi_step |
158
- | `prize_ratio_charity` | 5 | 83 | ratio |
159
-
160
- ### Correct-answer positions and numeric rank
161
-
162
- Both balanced, so neither slot priors nor magnitude priors are worth anything:
163
-
164
- | Axis | Distribution |
165
- | --- | --- |
166
- | Correct slot | 252 / 252 / 248 / 248 |
167
- | Answer numeric rank (smallest → largest) | 248 / 252 / 248 / 252 |
168
-
169
- The `slot × rank × depth` grid is filled at 15–16 items per cell.
170
 
171
- ## Shortcut audit
172
 
173
- Measured on this file (`python deeparithmark.py audit`); chance is 25.00 %:
174
-
175
- | Strategy that needs no arithmetic | Score |
176
- | --- | ---: |
177
- | pick largest value | 25.20 % |
178
- | pick smallest value | 24.80 % |
179
- | pick second smallest | 25.20 % |
180
- | pick the roundest number | 25.20 % |
181
- | pick longest / shortest text | 25.50 % / 24.00 % |
182
- | fixed slot 0 / 1 / 2 / 3 | 25.20 / 25.20 / 24.80 / 24.80 % |
183
- | char-3-gram classifier over option text | 25.47 % (max 28.00 % over 5 splits) |
184
-
185
- Nothing clears chance. This matters: an early build of this dataset scored gpt2
186
- **below** chance (21.1 %) because several decoys were structurally rounder than
187
- the answer, so a round-number prior was anti-correlated with correctness. The
188
- fix — all four options in the same roundness class — is documented in the design
189
- notes, and it is now a build constraint.
190
-
191
- ## Baselines
192
-
193
- Zero-shot, `acc_norm`, bf16, batch 32, `--max-context 1024` (ArithMark 3.0's
194
- protocol). ArithMark 3.0 numbers were produced on the same weights in the same
195
- run, not copied from a table.
196
 
197
  | Model | Params | Deep | soft | d2 | d3 | d4 | d5 | ArithMark-3 | Δ |
198
  | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
@@ -210,81 +145,79 @@ run, not copied from a table.
210
  | nanowhale-100m-base | 110 M | 21.30 % | 24.46 % | 15.6 | 22.8 | 23.6 | 23.2 | 26.50 % | −5.20 |
211
  | EleutherAI/pythia-70m | 70 M | 22.70 % | 24.34 % | 21.6 | 26.0 | 21.6 | 21.6 | 28.80 % | −6.10 |
212
 
213
- `d2`–`d5` are `acc_norm` per depth tier (250 items each, chance 25 %). `soft` is
214
- the mean probability a proper scoring rule gives the correct option after a
215
- softmax over the four continuation scores — the graded companion to hard
216
- accuracy. Δ is Deep Arithmark minus ArithMark 3.0 on the same weights.
217
  `Tev1-0.8B-experimental` is a post-trained derivative of `Qwen/Qwen3.5-0.8B`
218
  (identical parameter count), so that pair is base vs. post-trained on one
219
  architecture rather than two models. `gemma-3-270m` and `MobileLLM-R1-950M` are
220
  gated and have not been scored.
221
 
222
- **Read the baselines carefully.** Results split into three bands. Below ~350 M
223
- nothing clears chance and some scores sit significantly *below* it (nanowhale-100m
224
- 21.3 %, p = 0.007; LFM2.5-350M 20.0 %, p = 0.0002), which means a stable losing
225
- strategy rather than encoding arithmetic. **Qwen3-0.6B-Base is the first sub-1B
226
- base model to clear chance significantly: 29.20 %, p = 0.0027**, and the first to show a
227
- monotone depth decay (37.6 → 32.0 → 28.8 → 18.4 %), significant at both ends
228
- (depth 2 p = 1e-5, depth 5 p = 0.016). A rank-preference artifact cannot produce
229
- that shape, because option rank is balanced uniformly across depth. Above 1B the
230
- depth ladder is fully readable (Qwen2.5-1.5B-Instruct, 60.4 → 22.4 %).
231
-
232
- GPT-X3-150M (149 M) is the clearest illustration of what this dataset removes. It
233
- is the strongest 150M-class model here on ArithMark 3.0 — **49.60 %**, 11 points
234
- above SmolLM2-135M — and drops **24.4 points** to 25.20 % on this dataset. But it
235
- is not guessing uniformly: per-template accuracy runs from 3.0 % to 54.5 %
236
- (χ² = 63.2, df = 14, p < 1e-4), driven by a small-number prior (it picks the
237
- smallest option 34.8 % of the time, the largest 15.8 %) plus one stable wrong
238
- path per family in others. Those cancel to chance in aggregate, so the 25.20 % is
239
- a real reading of "cannot solve these", not noise — but treat per-template rows as
240
- preference, not skill. Full breakdown in the design notes, §6.
241
-
242
- The extreme case is granite-4.0-350m-base: **75.70 %** on ArithMark 3.0, the
243
- highest sub-1B result in the table, against 24.10 % here — a **51.6-point** drop.
244
- The measured form of "harder" is therefore not a small constant; it scales with
245
- how well a model exploits ArithMark 3.0's ±1 option structure.
246
 
247
  For a sub-1B leaderboard, report the per-depth curve rather than only the single
248
- number. The first real signal appears at the depth-2 rung, and at 1.5B-instruct
249
- the two solvable templates are `share_then_split` (86.4 %) and `time_clock_end`
250
- (85.9 %).
251
 
252
- ## Usage
253
-
254
- ```python
255
- from datasets import load_dataset
256
-
257
- ds = load_dataset("OPENGCM/DeepArithmark-1", split="test")
258
- item = ds[0]
259
- # score the four continuations of item["ctx"] by length-normalised
260
- # log-likelihood and take the argmax; compare it with int(item["label"])
261
- ```
262
-
263
- Or score Hugging Face models directly, including a side-by-side ArithMark 3.0
264
- comparison on the same weights:
265
-
266
- ```bash
267
- python bench_deeparithmark.py --compare --model HuggingFaceTB/SmolLM2-135M
268
- python bench_deeparithmark.py --model Qwen/Qwen2.5-0.5B --out benchmarks
269
- ```
270
-
271
- ArithMark 3.0's own script works unchanged:
272
-
273
- ```bash
274
- python bencharithmark-3.py --data-path data/deeparithmark-1.jsonl --model gpt2
275
- ```
276
-
277
- ## How it was generated
278
 
279
- `deeparithmark.py build` renders the file deterministically from seed
280
- `20261027`: 15 parameterised families, exact integer arithmetic, and a decoy
281
- pool per family of plausible wrong paths. `python deeparithmark.py selftest`
282
- re-checks every claim on this page — schema, unit consistency, option
283
- distinctness, step count vs depth, all three balance axes, and the shortcut
284
- audit.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
285
 
286
- The seed makes the *instances* reproducible and unseen; it does not make the
287
- template shapes novel. Do not train on this file.
 
 
288
 
289
  ## Limitations
290
 
@@ -294,27 +227,49 @@ template shapes novel. Do not train on this file.
294
  a gradient.
295
  - Single `test` split, no train split (evaluation set).
296
  - 15 templates only: fine-tuning on them would overfit.
 
 
297
  - Clock items keep all four options inside one half of the day, so no noon
298
  rollover is exercised.
299
- - Grade bands (`grades_4_5` … `grades_7_8`) describe the arithmetic, not a
300
- curriculum.
 
 
 
 
 
 
 
 
 
301
 
302
  ## Licence and attribution
303
 
304
- This dataset is released under the **Apache License 2.0**.
 
305
 
306
  **Upstream work.** Deep Arithmark follows [ArithMark 3.0](https://huggingface.co/datasets/AxiomicLabs/Arithmark-3.0)
307
- by AxiomicLabs, also Apache-2.0. What is reused is the *task contract*: one
308
- unfinished sentence, four continuations with a leading space, a single correct
309
- index, and scoring by length-normalised continuation log-likelihood — plus the
310
- JSONL schema, so ArithMark 3.0's own scorer runs on this file unchanged.
311
-
312
- **What is not reused.** No ArithMark 3.0 item appears here. Every prompt is
313
- generated from the parameterised families in `deeparithmark.py`,
314
- and the released file shares **0 identical contexts and 0 eight-word n-grams**
315
- with the 1,000 upstream items. The accompanying scorer is an independent
316
- implementation of the same protocol, not copied code.
317
-
318
- **Trademarks.** "ArithMark" and "AxiomicLabs" remain the property of their
319
- owner. Their use here is descriptive, to state which benchmark this one extends.
320
- It does not imply affiliation or endorsement — see the notice at the top.
 
 
 
 
 
 
 
 
 
 
 
20
  path: data/deeparithmark-1.jsonl
21
  ---
22
 
23
+ # Deep Arithmark 1
24
 
25
+ **1,000 arithmetic word problems that need 2–5 dependent operations.** Same
26
+ format as [ArithMark 3.0](https://huggingface.co/datasets/AxiomicLabs/Arithmark-3.0)
27
+ — finish the sentence, pick one of four continuations, chance 25 % — but the
28
+ wrong answers are no longer ±1 of the right one, so a magnitude guess does not
29
+ work.
30
+
31
+ > **Unofficial.** Deep Arithmark is an independent companion tier. It is not
32
+ > affiliated with, endorsed by, or published by AxiomicLabs, the authors of
33
+ > ArithMark 3.0, and it is not an official ArithMark release. It follows their
34
+ > task contract and credits them throughout; any errors are the author's. If you
35
+ > use this benchmark, cite ArithMark 3.0 as well.
36
+
37
+ | | |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38
  | --- | ---: |
39
  | Examples | 1,000 |
40
+ | Choices per example | 4 |
41
+ | Chance | 25 % |
 
 
42
  | Depth tiers | 2, 3, 4, 5 — 250 each |
43
+ | Templates / topic labels | 15 / 14 |
44
  | Mean prompt length | 29.2 words (max 43) |
45
+ | Language, source | English, synthetic (deterministic seed) |
 
 
 
 
46
 
47
+ ## Quick start
48
 
49
+ ```python
50
+ from datasets import load_dataset
51
+
52
+ ds = load_dataset("OPENGCM/DeepArithmark-1", split="test")
53
+ item = ds[0]
54
+ # score the four continuations of item["ctx"] by length-normalised
55
+ # log-likelihood, take the argmax, compare with int(item["label"])
56
+ ```
57
+
58
+ Score models directly, including an ArithMark 3.0 comparison on the same weights:
59
+
60
+ ```bash
61
+ python bench_deeparithmark.py --compare --model HuggingFaceTB/SmolLM2-135M
62
+ ```
63
+
64
+ ArithMark 3.0's own scorer runs on this file unchanged (verified against that
65
+ script's loader):
66
 
67
+ ```bash
68
+ python bencharithmark-3.py --data-path data/deeparithmark-1.jsonl --model gpt2
69
+ ```
70
+
71
+ ## Why it is harder
72
+
73
+ ArithMark 3.0's four continuations are **adjacent integers** (43 / 44 / 45 / 46).
74
+ Rejecting a wrong option only needs "the answer is about 44", so a magnitude
75
+ prior is sufficient and the arithmetic is optional. Deep Arithmark keeps the
76
+ contract and changes three things:
77
+
78
+ 1. **Depth.** Every item needs 2–5 dependent operations; later steps consume
79
+ earlier results. `depth` equals the length of `metadata.solution_steps`, and
80
+ the build fails otherwise.
81
+ 2. **Decoys with provenance.** No decoy is "answer ± 1". Each one is the exact
82
+ value of a specific plausible mistake — the state after step 2 instead of
83
+ step 5, the percentage on the wrong base, the divisor `n` instead of `n + 1`,
84
+ the dropped unit conversion. `metadata.option_sources` names the mistake per
85
+ ending, so error analysis can ask *which step* a model botches.
86
+ 3. **Balanced on three axes.** Correct slot, the answer's *numeric rank*, and the
87
+ *roundness class* of all four options are each balanced. "Pick the largest"
88
+ and "pick the roundest number" are both worth 25 %.
89
 
90
+ ## One item
 
 
91
 
92
  ```json
93
  {
 
117
  }
118
  ```
119
 
120
+ The answer is ` 26 crates.` — `420 + 260 − 210 − 40 = 430`, and `430 // 16 = 26`
121
+ full crates. `29` is what you get by forgetting the damaged units, `27` by
122
+ counting the loose units as a crate, `31` by adding them.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
123
 
124
+ ## Scores
125
 
126
+ Zero-shot, `acc_norm`, bf16, batch 32, `--max-context 1024`. ArithMark 3.0
127
+ numbers were produced on the same weights in the same run, not copied from a
128
+ table. `soft` is the mean probability of the correct option after a softmax over
129
+ the four scores; `d2`–`d5` are `acc_norm` per depth tier (250 items each, chance
130
+ 25 %). Δ is Deep Arithmark minus ArithMark 3.0.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
131
 
132
  | Model | Params | Deep | soft | d2 | d3 | d4 | d5 | ArithMark-3 | Δ |
133
  | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
 
145
  | nanowhale-100m-base | 110 M | 21.30 % | 24.46 % | 15.6 | 22.8 | 23.6 | 23.2 | 26.50 % | −5.20 |
146
  | EleutherAI/pythia-70m | 70 M | 22.70 % | 24.34 % | 21.6 | 26.0 | 21.6 | 21.6 | 28.80 % | −6.10 |
147
 
 
 
 
 
148
  `Tev1-0.8B-experimental` is a post-trained derivative of `Qwen/Qwen3.5-0.8B`
149
  (identical parameter count), so that pair is base vs. post-trained on one
150
  architecture rather than two models. `gemma-3-270m` and `MobileLLM-R1-950M` are
151
  gated and have not been scored.
152
 
153
+ **Three bands.** Below ~350 M nothing clears chance, and two models sit
154
+ significantly *below* it — nanowhale-100M (21.3 %, p = 0.007) and LFM2.5-350M
155
+ (20.0 %, p = 0.0002) — meaning a stable losing strategy rather than absent
156
+ arithmetic encoding. **Qwen3-0.6B-Base is the first sub-1B model to clear chance
157
+ significantly: 29.20 %, p = 0.0027.** Above 1B the depth ladder is fully readable.
158
+
159
+ **The one to watch is Qwen3-0.6B's depth curve:** 37.6 → 32.0 → 28.8 → **18.4 %**,
160
+ significant at both ends (depth 2 p = 1e-5, depth 5 p = 0.016). Being *below*
161
+ chance at depth 5 means it commits to a stable wrong path — it starts computing
162
+ and drops the last step — rather than guessing. Option rank is balanced uniformly
163
+ across depth, so no rank preference can produce that shape.
164
+
165
+ **Hardness scales with how well a model exploits ±1 options.** granite-4.0-350M
166
+ Base scores **75.70 %** on ArithMark 3.0, the highest sub-1B result here, and
167
+ 24.10 % on this dataset — a **51.6-point** drop, against GPT-X3-150M's 24.4.
168
+
169
+ **A chance score is not always guessing.** GPT-X3-150M scores 49.60 % on
170
+ ArithMark 3.0 and 25.20 % here, but its per-template accuracy spans 3.0 % to
171
+ 54.5 % (χ² = 63.2, df = 14, p < 1e-4): a small-number prior (picks the smallest
172
+ option 34.8 % of the time, the largest 15.8 %) plus one stable wrong path per
173
+ family. These cancel to chance in aggregate, so treat per-template rows as
174
+ preference, not partial skill. Worked through in the design notes, §6.
 
 
175
 
176
  For a sub-1B leaderboard, report the per-depth curve rather than only the single
177
+ number. At 1.5B-instruct the two solvable templates are `share_then_split`
178
+ (86.4 %) and `time_clock_end` (85.9 %).
 
179
 
180
+ ## Composition
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
181
 
182
+ | Depth | Examples | Typical item |
183
+ | ---: | ---: | --- |
184
+ | 2 | 250 | percent off a price, then a fee |
185
+ | 3 | 250 | unit conversion, add, subtract |
186
+ | 4 | 250 | receive / ship / damage / pack |
187
+ | 5 | 250 | two percentage days, then palletise |
188
+
189
+ | Template | Depth | Examples | Template | Depth | Examples |
190
+ | --- | ---: | ---: | --- | ---: | ---: |
191
+ | `money_discount_fee` | 2 | 122 | `schedule_sessions` | 3 | 47 |
192
+ | `time_clock_end` | 2 | 128 | `inventory_boxes` | 4 | 62 |
193
+ | `measure_cm_cut` | 3 | 33 | `mix_split` | 4 | 63 |
194
+ | `share_then_split` | 3 | 44 | `tank_fill` | 4 | 62 |
195
+ | `two_rate_pay` | 3 | 45 | `discount_tax_coupon_split` | 4 | 63 |
196
+ | `stock_percent_pack` | 3 | 48 | `salary_overtime_tax` | 5 | 84 |
197
+ | `trip_minutes` | 3 | 33 | `inventory_percent_chain` | 5 | 83 |
198
+ | | | | `prize_ratio_charity` | 5 | 83 |
199
+
200
+ Correct slot is 252 / 252 / 248 / 248, answer numeric rank (smallest → largest)
201
+ is 248 / 252 / 248 / 252, and the `slot × rank × depth` grid holds 15–16 items
202
+ per cell.
203
+
204
+ ### Shortcut audit
205
+
206
+ No strategy that skips the arithmetic beats chance (`python deeparithmark.py
207
+ audit`, chance 25.00 %):
208
+
209
+ | Strategy | Score |
210
+ | --- | ---: |
211
+ | pick largest / smallest value | 25.20 % / 24.80 % |
212
+ | pick the roundest number | 25.20 % |
213
+ | fixed slot 0 / 1 / 2 / 3 | 25.20 / 25.20 / 24.80 / 24.80 % |
214
+ | longest / shortest text | 25.50 % / 24.00 % |
215
+ | char-3-gram classifier over option text | 25.47 % |
216
 
217
+ This is a build constraint, not a happy accident. An early build scored gpt2
218
+ **below** chance (21.1 %) because several decoys were structurally rounder than
219
+ the answer, making a roundness prior anti-correlated with correctness. All four
220
+ options are now forced into the same roundness class.
221
 
222
  ## Limitations
223
 
 
227
  a gradient.
228
  - Single `test` split, no train split (evaluation set).
229
  - 15 templates only: fine-tuning on them would overfit.
230
+ - Rank balance is global, not per-template, so per-template rows read as
231
+ preference rather than skill.
232
  - Clock items keep all four options inside one half of the day, so no noon
233
  rollover is exercised.
234
+
235
+ ## How it was generated
236
+
237
+ `deeparithmark.py build` renders the file deterministically from seed `20261027`:
238
+ 15 parameterised families, exact integer arithmetic, one decoy pool per family.
239
+ `python deeparithmark.py selftest` re-checks every claim on this page — schema,
240
+ unit consistency, option distinctness, step count vs depth, all three balance
241
+ axes, and the audit above.
242
+
243
+ The seed makes the *instances* reproducible and unseen; it does not make the
244
+ template shapes novel. Do not train on this file.
245
 
246
  ## Licence and attribution
247
 
248
+ Released under the **Apache License 2.0**. Dataset SHA-256:
249
+ `8a54eb90b66b142b7fea422b34d851746d1cabc476dcba906c7f3ad5ddc66ee0`.
250
 
251
  **Upstream work.** Deep Arithmark follows [ArithMark 3.0](https://huggingface.co/datasets/AxiomicLabs/Arithmark-3.0)
252
+ by AxiomicLabs, also Apache-2.0. Reused: the *task contract* — one unfinished
253
+ sentence, four continuations with a leading space, a single correct index, and
254
+ scoring by length-normalised continuation log-likelihood — plus the JSONL
255
+ schema, so their scorer runs on this file unchanged.
256
+
257
+ **What is not reused.** No ArithMark 3.0 item appears here. Every prompt comes
258
+ from the parameterised families in `deeparithmark.py`, and the released file
259
+ shares **0 identical contexts and 0 eight-word n-grams** with the 1,000 upstream
260
+ items. The scorer is an independent implementation of the same protocol, not
261
+ copied code.
262
+
263
+ **Trademarks.** "ArithMark" and "AxiomicLabs" remain the property of their owner.
264
+ Their use here is descriptive, to state which benchmark this one extends. It does
265
+ not imply affiliation or endorsement — see the notice at the top.
266
+
267
+ ## Files
268
+
269
+ | File | What it is |
270
+ | --- | --- |
271
+ | `data/deeparithmark-1.jsonl` | The dataset — 1,000 items, one JSON object per line |
272
+ | `deeparithmark.py` | Deterministic generator, plus `selftest` and the shortcut `audit` |
273
+ | `bench_deeparithmark.py` | Scorer with per-depth and per-template breakdowns and an ArithMark 3.0 comparison |
274
+ | `DEEPARITHMARK_DESIGN.md` | Design rationale, fairness audit, measurement discussion |
275
+ | `benchmarks/deeparithmark_6models.json` | Raw results for the 2026 sub-1B sweep |