github/toon
1
0
Fork 0
mirror of https://github.com/voson-wang/toon.git synced 2026-01-29 15:24:10 +08:00
Code Issues Packages Projects Releases Wiki Activity

chore: split token efficiency benchmark into mixed/flat tracks

Browse Source
This commit is contained in:
Johann Schopplich
2025-11-06 22:17:18 +01:00
parent e22884308b
commit 54433de930
13 changed files with 567 additions and 1830 deletions
Download Patch File Download Diff File Expand all files Collapse all files

227
benchmarks/results/retrieval-accuracy.md
View File

@@ -1,108 +1,153 @@
Benchmarks test LLM comprehension across different input formats using 154 data retrieval questions on 4 models.
Benchmarks test LLM comprehension across different input formats using 201 data retrieval questions on 4 models.
<details>
<summary><strong>View Dataset Catalog</strong></summary>
#### Dataset Catalog
| Dataset | Rows | Structure | CSV Support | Eligibility |
| ------- | ---- | --------- | ----------- | ----------- |
| Uniform employee records | 100 | uniform | ✓ | 100% |
| E-commerce orders with nested structures | 50 | nested | ✗ | 33% |
| Time-series analytics data | 60 | uniform | ✓ | 100% |
| Top 100 GitHub repositories | 100 | uniform | ✓ | 100% |
| Semi-uniform event logs | 75 | semi-uniform | ✗ | 50% |
| Deeply nested configuration | 11 | deep | ✗ | 0% |
**Structure classes:**
- **uniform**: All objects have identical fields with primitive values
- **semi-uniform**: Mix of uniform and non-uniform structures
- **nested**: Objects with nested structures (nested objects or arrays)
- **deep**: Highly nested with minimal tabular eligibility
**CSV Support:** ✓ (supported), ✗ (not supported – would require lossy flattening)
**Eligibility:** Percentage of arrays that qualify for TOON's tabular format (uniform objects with primitive values)
</details>
#### Efficiency Ranking (Accuracy per 1K Tokens)
Each format's overall performance, balancing accuracy against token cost:
```
toon ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ 15.0 │ 70.1% acc │ 4,678 tokens
csv ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░ 14.3 │ 67.7% acc │ 4,745 tokens
json-compact ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░ 11.0 │ 65.3% acc │ 5,925 tokens
yaml ▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░ 9.4 │ 66.7% acc │ 7,091 tokens
json-pretty ▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░ 7.5 │ 65.4% acc │ 8,713 tokens
xml ▓▓▓▓▓▓▓▓▓░░░░░░░░░░░ 6.8 │ 67.2% acc │ 9,944 tokens
TOON ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ 15.6 │ 68.7% acc │ 4,389 tokens
CSV ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ 15.3 │ 62.3% acc │ 4,080 tokens
JSON compact ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░ 13.5 │ 67.2% acc │ 4,982 tokens
YAML ▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░ 11.2 │ 66.7% acc │ 5,976 tokens
JSON ▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░ 9.0 │ 65.7% acc │ 7,260 tokens
XML ▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░ 8.1 │ 66.8% acc │ 8,251 tokens
```
TOON achieves **70.1%** accuracy (vs JSON's 65.4%) while using **46.3% fewer tokens**.
TOON achieves **68.7%** accuracy (vs JSON's 65.7%) while using **39.5% fewer tokens**.
#### Per-Model Accuracy
Accuracy across **4 LLMs** on 154 data retrieval questions:
Accuracy across 4 LLMs on 201 data retrieval questions:
```
gpt-5-nano
→ TOON ███████████████████░ 96.1% (148/154)
CSV ██████████████████░░ 91.6% (141/154)
YAML ██████████████████░░ 91.6% (141/154)
JSON compact ██████████████████░░ 91.6% (141/154)
XML █████████████████░░░ 87.0% (134/154)
JSON █████████████████░░░ 86.4% (133/154)
→ TOON ██████████████████░░ 88.6% (178/201)
JSON compact ██████████████████░░ 88.1% (177/201)
CSV ██████████████████░░ 88.0% (88/100)
YAML █████████████████░░░ 84.6% (170/201)
XML ████████████████░░░░ 81.6% (164/201)
JSON ████████████████░░░░ 80.1% (161/201)
claude-haiku-4-5-20251001
JSON ██████████░░░░░░░░░░ 50.0% (77/154)
YAML ██████████░░░░░░░░░░ 49.4% (76/154)
→ TOON ██████████░░░░░░░░░░ 48.7% (75/154)
XML ██████████░░░░░░░░░░ 48.1% (74/154)
CSV █████████░░░░░░░░░░░ 47.4% (73/154)
JSON compact █████████░░░░░░░░░░░ 44.2% (68/154)
YAML ██████████░░░░░░░░░░ 52.2% (105/201)
→ TOON ██████████░░░░░░░░░░ 50.7% (102/201)
JSON ██████████░░░░░░░░░░ 50.2% (101/201)
JSON compact ██████████░░░░░░░░░░ 49.8% (100/201)
XML ██████████░░░░░░░░░░ 49.3% (99/201)
CSV ████████░░░░░░░░░░░░ 39.0% (39/100)
gemini-2.5-flash
CSV ██████████████████░░ 87.7% (135/154)
XML ██████████████████░░ 87.7% (135/154)
→ TOON █████████████████░░░ 86.4% (133/154)
YAML ████████████████░░░░ 79.9% (123/154)
JSON compact ████████████████░░░░ 79.9% (123/154)
JSON ███████████████░░░░░ 76.6% (118/154)
XML █████████████████░░░ 86.1% (173/201)
→ TOON █████████████████░░░ 84.1% (169/201)
CSV ████████████████░░░░ 82.0% (82/100)
JSON compact ████████████████░░░░ 81.1% (163/201)
YAML ████████████████░░░░ 81.1% (163/201)
JSON ████████████████░░░░ 81.1% (163/201)
grok-4-fast-non-reasoning
→ TOON ██████████░░░░░░░░░░ 49.4% (76/154)
JSON ██████████░░░░░░░░░░ 48.7% (75/154)
XML █████████░░░░░░░░░░░ 46.1% (71/154)
YAML █████████░░░░░░░░░░░ 46.1% (71/154)
JSON compact █████████░░░░░░░░░░░ 45.5% (70/154)
CSV █████████░░░░░░░░░░░ 44.2% (68/154)
→ TOON ██████████░░░░░░░░░░ 51.2% (103/201)
JSON ██████████░░░░░░░░░░ 51.2% (103/201)
XML ██████████░░░░░░░░░░ 50.2% (101/201)
JSON compact ██████████░░░░░░░░░░ 49.8% (100/201)
YAML ██████████░░░░░░░░░░ 48.8% (98/201)
CSV ████████░░░░░░░░░░░░ 40.0% (40/100)
```
**Key tradeoff:** TOON achieves **70.1% accuracy** (vs JSON's 65.4%) while using **46.3% fewer tokens** on these datasets.
**Key tradeoff:** TOON achieves **68.7% accuracy** (vs JSON's 65.7%) while using **39.5% fewer tokens** on these datasets.
<details>
<summary><strong>Performance by dataset and model</strong></summary>
#### Performance by Dataset
##### Uniform employee records (TOON optimal format)
##### Uniform employee records
| Format | Accuracy | Tokens | Correct/Total |
| ------ | -------- | ------ | ------------- |
| `csv` | 65.5% | 2,337 | 131/200 |
| `toon` | 67.5% | 2,483 | 135/200 |
| `json-compact` | 65.5% | 3,943 | 131/200 |
| `yaml` | 68.5% | 4,969 | 137/200 |
| `xml` | 69.5% | 7,314 | 139/200 |
| `json-pretty` | 64.5% | 6,347 | 129/200 |
| `toon` | 65.6% | 2,483 | 105/160 |
| `csv` | 62.5% | 2,337 | 100/160 |
| `json-compact` | 66.3% | 3,943 | 106/160 |
| `yaml` | 63.7% | 4,969 | 102/160 |
| `xml` | 67.5% | 7,314 | 108/160 |
| `json-pretty` | 62.5% | 6,347 | 100/160 |
##### E-commerce orders with nested structures
| Format | Accuracy | Tokens | Correct/Total |
| ------ | -------- | ------ | ------------- |
| `toon` | 78.8% | 5,967 | 126/160 |
| `csv` | 76.3% | 6,735 | 122/160 |
| `json-compact` | 70.6% | 5,962 | 113/160 |
| `yaml` | 72.5% | 7,328 | 116/160 |
| `json-pretty` | 76.9% | 9,694 | 123/160 |
| `xml` | 73.1% | 10,992 | 117/160 |
| `toon` | 75.6% | 7,197 | 121/160 |
| `json-compact` | 70.6% | 6,784 | 113/160 |
| `yaml` | 71.9% | 8,334 | 115/160 |
| `json-pretty` | 68.8% | 10,700 | 110/160 |
| `xml` | 71.9% | 12,013 | 115/160 |
##### Time-series analytics data
| Format | Accuracy | Tokens | Correct/Total |
| ------ | -------- | ------ | ------------- |
| `toon` | 68.4% | 1,515 | 93/136 |
| `csv` | 65.4% | 1,393 | 89/136 |
| `json-compact` | 64.7% | 2,341 | 88/136 |
| `yaml` | 66.2% | 2,938 | 90/136 |
| `json-pretty` | 64.7% | 3,665 | 88/136 |
| `xml` | 66.9% | 4,376 | 91/136 |
| `csv` | 63.8% | 1,391 | 74/116 |
| `toon` | 66.4% | 1,513 | 77/116 |
| `json-compact` | 61.2% | 2,339 | 71/116 |
| `yaml` | 65.5% | 2,936 | 76/116 |
| `json-pretty` | 64.7% | 3,663 | 75/116 |
| `xml` | 65.5% | 4,374 | 76/116 |
##### Top 100 GitHub repositories
| Format | Accuracy | Tokens | Correct/Total |
| ------ | -------- | ------ | ------------- |
| `toon` | 65.0% | 8,745 | 78/120 |
| `csv` | 62.5% | 8,513 | 75/120 |
| `json-compact` | 58.3% | 11,455 | 70/120 |
| `yaml` | 56.7% | 13,129 | 68/120 |
| `xml` | 55.8% | 17,095 | 67/120 |
| `json-pretty` | 52.5% | 15,145 | 63/120 |
| `toon` | 63.7% | 8,745 | 79/124 |
| `csv` | 60.5% | 8,513 | 75/124 |
| `json-compact` | 56.5% | 11,455 | 70/124 |
| `yaml` | 53.2% | 13,129 | 66/124 |
| `json-pretty` | 53.2% | 15,145 | 66/124 |
| `xml` | 53.2% | 17,095 | 66/124 |
##### Semi-uniform event logs
| Format | Accuracy | Tokens | Correct/Total |
| ------ | -------- | ------ | ------------- |
| `json-compact` | 55.0% | 4,809 | 66/120 |
| `yaml` | 52.5% | 5,814 | 63/120 |
| `json-pretty` | 52.5% | 6,784 | 63/120 |
| `toon` | 45.8% | 5,764 | 55/120 |
| `xml` | 50.8% | 7,699 | 61/120 |
##### Deeply nested configuration
| Format | Accuracy | Tokens | Correct/Total |
| ------ | -------- | ------ | ------------- |
| `json-compact` | 91.9% | 564 | 114/124 |
| `toon` | 92.7% | 631 | 115/124 |
| `yaml` | 91.9% | 673 | 114/124 |
| `json-pretty` | 91.9% | 919 | 114/124 |
| `xml` | 89.5% | 1,008 | 111/124 |
#### Performance by Model
@@ -110,45 +155,45 @@ grok-4-fast-non-reasoning
| Format | Accuracy | Correct/Total |
| ------ | -------- | ------------- |
| `toon` | 96.1% | 148/154 |
| `csv` | 91.6% | 141/154 |
| `yaml` | 91.6% | 141/154 |
| `json-compact` | 91.6% | 141/154 |
| `xml` | 87.0% | 134/154 |
| `json-pretty` | 86.4% | 133/154 |
| `toon` | 88.6% | 178/201 |
| `json-compact` | 88.1% | 177/201 |
| `csv` | 88.0% | 88/100 |
| `yaml` | 84.6% | 170/201 |
| `xml` | 81.6% | 164/201 |
| `json-pretty` | 80.1% | 161/201 |
##### claude-haiku-4-5-20251001
| Format | Accuracy | Correct/Total |
| ------ | -------- | ------------- |
| `json-pretty` | 50.0% | 77/154 |
| `yaml` | 49.4% | 76/154 |
| `toon` | 48.7% | 75/154 |
| `xml` | 48.1% | 74/154 |
| `csv` | 47.4% | 73/154 |
| `json-compact` | 44.2% | 68/154 |
| `yaml` | 52.2% | 105/201 |
| `toon` | 50.7% | 102/201 |
| `json-pretty` | 50.2% | 101/201 |
| `json-compact` | 49.8% | 100/201 |
| `xml` | 49.3% | 99/201 |
| `csv` | 39.0% | 39/100 |
##### gemini-2.5-flash
| Format | Accuracy | Correct/Total |
| ------ | -------- | ------------- |
| `csv` | 87.7% | 135/154 |
| `xml` | 87.7% | 135/154 |
| `toon` | 86.4% | 133/154 |
| `yaml` | 79.9% | 123/154 |
| `json-compact` | 79.9% | 123/154 |
| `json-pretty` | 76.6% | 118/154 |
| `xml` | 86.1% | 173/201 |
| `toon` | 84.1% | 169/201 |
| `csv` | 82.0% | 82/100 |
| `json-compact` | 81.1% | 163/201 |
| `yaml` | 81.1% | 163/201 |
| `json-pretty` | 81.1% | 163/201 |
##### grok-4-fast-non-reasoning
| Format | Accuracy | Correct/Total |
| ------ | -------- | ------------- |
| `toon` | 49.4% | 76/154 |
| `json-pretty` | 48.7% | 75/154 |
| `xml` | 46.1% | 71/154 |
| `yaml` | 46.1% | 71/154 |
| `json-compact` | 45.5% | 70/154 |
| `csv` | 44.2% | 68/154 |
| `toon` | 51.2% | 103/201 |
| `json-pretty` | 51.2% | 103/201 |
| `xml` | 50.2% | 101/201 |
| `json-compact` | 49.8% | 100/201 |
| `yaml` | 48.8% | 98/201 |
| `csv` | 40.0% | 40/100 |
</details>
@@ -161,34 +206,36 @@ This benchmark tests **LLM comprehension and data retrieval accuracy** across di
#### Datasets Tested
Four datasets designed to test different structural patterns (all contain arrays of uniform objects, TOON's optimal format):
Six datasets designed to test different structural patterns:
1. **Tabular** (100 employee records): Uniform objects with identical fields – optimal for TOON's tabular format.
2. **Nested** (50 e-commerce orders): Complex structures with nested customer objects and item arrays.
3. **Analytics** (60 days of metrics): Time-series data with dates and numeric values.
4. **GitHub** (100 repositories): Real-world data from top GitHub repos by stars.
5. **Event Logs** (75 logs): Semi-uniform data with ~50% flat logs and ~50% with nested error objects.
6. **Nested Config** (1 configuration): Deeply nested configuration with minimal tabular eligibility.
#### Question Types
154 questions are generated dynamically across three categories:
201 questions are generated dynamically across three categories:
- **Field retrieval (40%)**: Direct value lookups or values that can be read straight off a record (including booleans and simple counts such as array lengths)
- **Field retrieval (36%)**: Direct value lookups or values that can be read straight off a record (including booleans and simple counts such as array lengths)
- Example: "What is Alice's salary?" → `75000`
- Example: "How many items are in order ORD-0042?" → `3`
- Example: "What is the customer name for order ORD-0042?" → `John Doe`
- **Aggregation (32%)**: Dataset-level totals and averages plus single-condition filters (counts, sums, min/max comparisons)
- **Aggregation (37%)**: Dataset-level totals and averages plus single-condition filters (counts, sums, min/max comparisons)
- Example: "How many employees work in Engineering?" → `17`
- Example: "What is the total revenue across all orders?" → `45123.50`
- Example: "How many employees have salary > 80000?" → `23`
- **Filtering (28%)**: Multi-condition queries requiring compound logic (AND constraints across fields)
- **Filtering (27%)**: Multi-condition queries requiring compound logic (AND constraints across fields)
- Example: "How many employees in Sales have salary > 80000?" → `5`
- Example: "How many active employees have more than 10 years of experience?" → `8`
#### Evaluation Process
1. **Format conversion**: Each dataset is converted to all 6 formats (TOON, CSV, XML, YAML, JSON, JSON compact).
1. **Format conversion**: Each dataset is converted to all 6 formats (TOON, JSON compact, XML, YAML, JSON, CSV).
2. **Query LLM**: Each model receives formatted data + question in a prompt and extracts the answer.
3. **Validate with LLM-as-judge**: `gpt-5-nano` validates if the answer is semantically correct (e.g., `50000` = `$50,000`, `Engineering` = `engineering`, `2025-01-01` = `January 1, 2025`).
@@ -197,6 +244,6 @@ Four datasets designed to test different structural patterns (all contain arrays
- **Models tested**: `gpt-5-nano`, `claude-haiku-4-5-20251001`, `gemini-2.5-flash`, `grok-4-fast-non-reasoning`
- **Token counting**: Using `gpt-tokenizer` with `o200k_base` encoding (GPT-5 tokenizer)
- **Temperature**: Not set (models use their defaults)
- **Total evaluations**: 154 questions × 6 formats × 4 models = 3,696 LLM calls
- **Total evaluations**: 201 questions × 6 formats × 4 models = 4,824 LLM calls
</details>