COINjecture NP-Solutions Dataset

Dataset Description

This dataset contains real-time blockchain data from the COINjecture Network, a proof-of-useful-work (PoUW) blockchain that uses NP-complete problems for consensus. This is a unified, continuous dataset that includes all problem types (SubsetSum, SAT, TSP, Custom) and consensus blocks in a single repository for comprehensive analysis.

Dataset Summary

The COINjecture Network is a blockchain that replaces traditional proof-of-work mining with solving useful computational problems. This unified dataset captures:

• Problem Submissions: NP-complete problems (SubsetSum, SAT, TSP, Custom) submitted to the network
• Solution Submissions: Solutions to problems with verification metrics
• Consensus Blocks: Complete block data including transactions, PoUW metrics, and energy measurements

Records are produced by network nodes (see coinject_huggingface::DatasetRecord in this repo) and uploaded as JSONL to the Hub. They are not the same artifact as other exports (e.g. API index tables); use this dataset for raw node-emitted training and research corpora.

All problem types are stored in a single continuous dataset (COINjecture/NP-Solutions) to enable cross-problem-type analysis and unified research workflows.

Supported Tasks

• Research: Study of NP-complete problem solving performance
• Energy Analysis: Energy consumption patterns in computational problem solving
• Blockchain Analytics: Consensus mechanism performance and transparency metrics
• Machine Learning: Training models on problem-solution pairs

Languages

English (problem descriptions and metadata)

Explorer-style layout (reference)

The JSONL rows are the source of truth; the layout below is the recommended human-readable presentation for explorers, dashboards, and docs. Times like 15s ago are computed from timestamp (Unix seconds) relative to “now” when rendering.

Example — SAT (consensus / mined block)

Block #123525
SAT
15s ago

Problem:
Satisfy 78 clauses with 26 variables

Solution:
Satisfying assignment found

Solver: 74446bf9...d77e15

Reward (BEANS)
124,324,271

Work (bits)
12.432

Asymmetry
5725.40×

Quality
1.000

Δt: 28.6 ms solve / 0.01 ms verify
Est. energy: 2.9 J

Example — SubsetSum (consensus / mined block)

Block #123524
SubsetSum
20s ago

Problem:
Find subset summing to 3938
Values: [55, 683, 630, 222, 651, 376, 332, 38, 827, 191, 292, 485, 453, 744, 403, 283, 717, 823, 350, 55, 928, 967, 995, 384, 354, 979, 733, 488, 882, 708, 67, 309, 751, 831]

Solution:
Indices 9, 10, 18, 19, 23, 29, 30, 31, 32, 33 → Sum: 3938

Solver: 74446bf9...d77e15

Reward (BEANS)
110,826,140

Work (bits)
11.083

Asymmetry
2845.00×

Quality
1.000

Δt: 2.8 ms solve / 0.00 ms verify
Est. energy: (from total_energy_joules when present)

Precomputed explorer_card

Current nodes set explorer_card on each emitted row to the same layout as below (UTC time line). For custom viewers you can print record["explorer_card"] directly, or rebuild from fields using the Python helper in this README.

Line-by-line mapping (JSONL → display)

Display	JSON fields / rule
Block #…	block_height
Type line (SAT, SubsetSum, …)	problem_type
Relative time	timestamp vs viewer clock (e.g. format_relative(timestamp))
Problem:	Derived from problem_data by type (see below)
Solution:	Derived from solution_data + problem_data (see below)
Solver:	solver or submitter (hex); show as first8...last6 for privacy
Reward (BEANS)	bounty (string u128) or formatted integer — native reward units on the network
Work (bits)	work_score when set; format with fixed decimals (e.g. 3)
Asymmetry	time_asymmetry (solve/verify time ratio); suffix ×
Quality	solution_quality when set (0–1 scale)
Δt:	solve_time_us, verify_time_us → ms: solve_time_us / 1000, verify_time_us / 1000
Est. energy	total_energy_joules (or sum of solve/verify energy fields) with one decimal and J

SAT — Problem line: From problem_data.clauses length and problem_data.variables (or equivalent):
Satisfy {n_clauses} clauses with {n_vars} variables.

SAT — Solution line: If solution_data.assignments exists: “Satisfying assignment found” (or list assignment preview for research dumps).

SubsetSum — Problem line: Find subset summing to {problem_data.target} plus Values: {problem_data.numbers} (truncate with “…” if extremely long).

SubsetSum — Solution line: Indices {comma-separated} → Sum: {target} where indices are solution_data.indices and target is problem_data.target (recompute sum for verification in tooling).

TSP / Custom: Use the same block header; problem/solution lines should summarize problem_data / solution_data (tour length, custom label) — extend the same pattern.

Optional: Python sketch

from __future__ import annotations

import time
from typing import Any, Mapping


def _rel_ago(ts: int) -> str:
    s = max(0, int(time.time()) - int(ts))
    if s < 60:
        return f"{s}s ago"
    if s < 3600:
        return f"{s // 60}m ago"
    return f"{s // 3600}h ago"


def _addr_short(hex64: str | None) -> str | None:
    if not hex64 or len(hex64) < 16:
        return hex64
    return f"{hex64[:8]}...{hex64[-6:]}"


def _fmt_int_string(s: str | None) -> str:
    if not s:
        return "—"
    try:
        return f"{int(s):,}"
    except ValueError:
        return s


def problem_line(pt: str, pd: Mapping[str, Any]) -> str:
    if pt == "SAT":
        n_c = len(pd.get("clauses") or [])
        n_v = int(pd.get("variables") or 0)
        return f"Satisfy {n_c} clauses with {n_v} variables"
    if pt == "SubsetSum":
        nums = pd.get("numbers") or []
        tgt = pd.get("target")
        return f"Find subset summing to {tgt}\nValues: {nums}"
    return str(pd)


def solution_line(pt: str, pd: Mapping[str, Any], sd: Mapping[str, Any] | None) -> str:
    if sd is None:
        return "—"
    if pt == "SAT":
        return "Satisfying assignment found"
    if pt == "SubsetSum":
        idx = sd.get("indices") or []
        tgt = pd.get("target")
        return f"Indices {', '.join(str(i) for i in idx)} → Sum: {tgt}"
    return str(sd)


def format_block_card(r: Mapping[str, Any]) -> str:
    pt = r.get("problem_type") or "?"
    pd = r.get("problem_data") or {}
    sd = r.get("solution_data")
    ws = r.get("work_score")
    ta = r.get("time_asymmetry")
    q = r.get("solution_quality")
    su = r.get("solve_time_us") or 0
    vu = r.get("verify_time_us") or 0
    ej = r.get("total_energy_joules")

    lines = [
        f"Block #{r.get('block_height', '?')}",
        str(pt),
        _rel_ago(int(r.get("timestamp") or 0)),
        "",
        "Problem:",
        problem_line(pt, pd),
        "",
        "Solution:",
        solution_line(pt, pd, sd),
        "",
        f"Solver: {_addr_short(r.get('solver') or r.get('submitter'))}",
        "",
        "Reward (BEANS)",
        _fmt_int_string(r.get("bounty")),
        "",
        "Work (bits)",
        f"{ws:.3f}" if isinstance(ws, (int, float)) else "—",
        "",
        "Asymmetry",
        f"{ta:.2f}×" if isinstance(ta, (int, float)) else "—",
        "",
        "Quality",
        f"{q:.3f}" if isinstance(q, (int, float)) else "—",
        "",
        f"Δt: {su / 1000:.1f} ms solve / {vu / 1000:.2f} ms verify",
    ]
    if isinstance(ej, (int, float)):
        lines.append(f"Est. energy: {ej:.1f} J")
    return "\n".join(lines)

Dataset Structure

Data Instances

Each record in the dataset represents either:

1. A problem submission (when a problem is submitted to the network)
2. A solution submission (when a solution is verified)
3. A consensus block (complete block data with all transactions)

Data Fields

Field	Type	Description
PRIMARY CONTENT
problem_id	string	Unique identifier for the problem
problem_type	string	Type of problem: "SubsetSum", "SAT", "TSP", "Custom", or "Private"
problem_data	object	Complete problem data. Always a JSON object in rows emitted by current nodes; very old JSONL may have used a string (double-encoded JSON), which breaks automatic schema inference across files.
solution_data	object (optional)	Solution data with normalized structure. Same object-vs-string caveat as problem_data for legacy shards.
explorer_card	string	Preformatted explorer-style card (multi-line text). Uses absolute UTC from timestamp in the card (not “Ns ago”). Omitted or empty on legacy JSONL without this field.
IDENTIFIERS
block_height	int64	Block height when the record was created
timestamp	int64	Unix timestamp (consensus rows: block header time; marketplace rows may use ingest time — see metrics_source)
submitter	string (optional)	Address of the problem submitter (hex encoded)
solver	string (optional)	Address of the solution solver (hex encoded)
PERFORMANCE METRICS
problem_complexity	float64	Complexity score of the problem
bounty	string	Bounty amount in native tokens (serialized as string to avoid JSON precision loss)
work_score	float64 (optional)	Work score calculated for the solution
solution_quality	float64 (optional)	Quality score of the solution
ASYMMETRY METRICS
time_asymmetry	float64 (optional)	Ratio of solve_time / verify_time
space_asymmetry	float64 (optional)	Memory asymmetry metric
energy_asymmetry	float64 (optional)	Energy asymmetry ratio
ENERGY MEASUREMENTS
solve_energy_joules	float64 (optional)	Energy consumed during solving (joules)
verify_energy_joules	float64 (optional)	Energy consumed during verification (joules)
total_energy_joules	float64 (optional)	Total energy consumption (joules)
energy_per_operation	float64 (optional)	Energy per operation estimate
energy_efficiency	float64 (optional)	Energy efficiency metric
TIMING (consensus / detailed rows)
solve_time_us	uint64 (optional)	Solve duration in microseconds (→ ms in explorer)
verify_time_us	uint64 (optional)	Verify duration in microseconds
mining_attempts	uint64 or null	Present on every JSONL row (null when not applicable). For consensus blocks the node sets this to the header nonce (on-chain proxy for search effort). Omitting the key across some shards caused Hugging Face Data Studio column mismatches.
MINING / CONSENSUS
difficulty_target	uint32 (optional)	Minimum leading zero bits in block hash (node PoW setting)
nonce	uint64 (optional)	Winning header nonce
METADATA
status	string	Status: "Pending", "Solved", "Mined", "Validated", etc.
submission_mode	string	Submission mode: "public", "private", or "mining"
energy_measurement_method	string	Method used: "rapl", "powermetrics", or "estimate"
DATA PROVENANCE
metrics_source	string	Source of metrics: "block_header_actual", "measured_marketplace", "estimated", or "not_applicable"
measurement_confidence	string	Confidence level: "high" (from header), "medium" (proxy/measured), "low" (estimate), or "not_applicable"
data_version	string	Dataset schema version (e.g. v3.1 — see huggingface/src/metrics.rs)

Consensus and marketplace paths may populate additional optional fields (timing, memory, energy, network, mining, hardware, economics). The full schema is DatasetRecord in huggingface/src/client.rs.

Solution Data Structure

Solutions are normalized to a consistent structure to avoid schema conflicts:

{
  "type": "SubsetSum" | "SAT" | "TSP" | "Custom",
  "data": <normalized data>
}

• SubsetSum: data is an array of indices (numbers)
• SAT: data is an array of 0/1 values (normalized from booleans)
• TSP: data is an array representing the tour (numbers)
• Custom: data is a base64-encoded string

Problem Data Structure

For consensus blocks, problem_data contains comprehensive block information:

{
  "height": <block_height>,
  "miner": <miner_address>,
  "transactions": [...],
  "solution_reveal": {
    "problem": {...},
    "solution": {
      "type": "...",
      "data": [...]
    },
    "commitment_hash": "...",
    "problem_hash": "..."
  },
  "solve_time_us": <time_in_microseconds>,
  "verify_time_us": <time_in_microseconds>,
  "energy_estimate_joules": <energy>,
  ...
}

Dataset Creation

Source Data

Data is collected in real-time from running COINjecture Network nodes. Each node pushes records to this dataset when:

• A problem is submitted via transaction
• A solution is submitted and verified
• A consensus block is mined or validated

Data Collection Process

1. Problem Submission: When a problem transaction is processed, a record is created with problem data
2. Solution Submission: When a solution is verified, metrics are calculated and a record is created
3. Consensus Blocks: Complete block data is recorded for transparency and analysis

Data Preprocessing

• Solutions are normalized to consistent schema (see Solution Data Structure)
• Energy measurements use multiple methods (RAPL, powermetrics, or estimation)
• Addresses are hex-encoded for consistency
• Timestamps are Unix epoch seconds
• Large integers (u128) are serialized as strings to avoid JSON precision loss
• All problem types are unified in a single continuous dataset for cross-problem analysis

Normalizing legacy JSONL (Hub viewer / CastError)

Older JSONL omitted optional keys on some rows. That yields different Arrow column sets across shards and breaks Data Studio with CastError ("column names don't match"). Current nodes emit the full key set on every line (null when unknown).

To batch-fix existing data/*.jsonl files locally (same key names as DatasetRecord in huggingface/src/client.rs; unknown top-level keys are dropped):

python3 scripts/hf_np_solutions_normalize_jsonl.py --in data/data_1775801281.jsonl --out data/data_1775801281.norm.jsonl

Then replace the originals on the Hub (e.g. hf upload or Hub API commits). For thousands of files, run in a loop or job runner and commit in batches. If you change DatasetRecord, update RECORD_KEYS in scripts/hf_np_solutions_jsonl_common.py to match.

Automated Hub pass (download → normalize → upload one commit per file; needs pip install huggingface_hub and a write token):

export HF_TOKEN=hf_...
python3 scripts/hf_np_solutions_batch_normalize_hub.py --dry-run
python3 scripts/hf_np_solutions_batch_normalize_hub.py --limit 5
python3 scripts/hf_np_solutions_batch_normalize_hub.py --sleep 1.0

Use --start-after data/data_<timestamp>.jsonl to resume. Full runs create thousands of commits; prefer a VM, tune --sleep, or fork the script to batch multiple files per create_commit if you hit rate limits.

Dataset Statistics

• Total Records: Growing in real-time (unified dataset with all problem types)
• Update Frequency: Real-time (buffered, flushed when 10 total records accumulated across all problem types)
• Data Format: JSONL (newline-delimited JSON)
• Storage Location: /data/ directory in the repository
• Problem Types: SubsetSum, SAT, TSP, Custom, Private (all in one dataset)
• Data Quality: v3.1 institutional-grade records when emitted by current nodes (block header and extended metrics where available)

Considerations for Using the Data

Ethical Considerations

• All data is from public blockchain transactions
• Addresses are included only if explicitly enabled (privacy option)
• No personally identifiable information is collected

Licensing

This dataset is released under the MIT License.

Citation Information

If you use this dataset in your research, please cite:

@dataset{coinjecture_np_solutions,
  title={COINjecture NP-Solutions Dataset},
  author={COINjecture Network},
  year={2024},
  url={https://huggingface.co/datasets/COINjecture/NP-Solutions}
}

Dataset Access

Using Hugging Face Datasets

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("COINjecture/NP-Solutions", split="train")

# Access records
for record in dataset:
    print(record["problem_id"])
    print(record["problem_data"])

Direct File Access

The raw JSONL files are available in the /data/ directory:

• Files are named data_<timestamp>.jsonl
• Each line is a complete JSON record
• Files can be processed with standard JSONL tools

API Access

The dataset is accessible via the Hugging Face API:

• Dataset viewer: https://huggingface.co/datasets/COINjecture/NP-Solutions
• API endpoint: https://huggingface.co/api/datasets/COINjecture/NP-Solutions

Uploading with the Hugging Face CLI

For manual pushes (exports, Parquet/JSONL, README updates), use the hf CLI. Install one of these (pick what works on your machine):

# Option A — standalone installer (macOS / Linux; adds hf to your PATH)
curl -LsSf https://hf.co/cli/install.sh | bash

# Option B — Homebrew (if the formula is available on your Mac)
brew install hf

# Option C — Python (hf ships with huggingface_hub; ensure the install’s bin is on PATH)
python3 -m pip install -U "huggingface_hub"

Then authenticate and upload (example: dataset card only from this repo):

hf auth login   # or: export HF_TOKEN=...  (never commit tokens)

cd /path/to/COINjecture2.0-main
hf upload COINjecture/NP-Solutions huggingface/README.md --repo-type=dataset \
  --commit-message "Dataset card: viewer YAML + schema notes"

# Or upload everything in the current directory tree:
hf upload COINjecture/NP-Solutions . --repo-type=dataset

Set HF_DATASET_NAME / --hf-dataset-name to this repo’s Hub id (COINjecture/NP-Solutions). Hyphen vs underscore are different Hub repositories if both exist.

Additional Information

Energy Measurement Methods

• RAPL (Linux): Intel/AMD Running Average Power Limit counters
• powermetrics (macOS): macOS powermetrics tool
• estimate: CPU TDP-based estimation (fallback, works everywhere)

Problem Types

1. SubsetSum: Find a subset of numbers that sum to a target
2. SAT: Boolean satisfiability problem
3. TSP: Traveling Salesman Problem
4. Custom: Arbitrary problem data (base64 encoded)

Performance Metrics

• Time Asymmetry: Measures how much harder solving is than verifying
• Space Asymmetry: Memory usage differences
• Energy Asymmetry: Energy consumption differences
• Energy Efficiency: Work performed per unit of energy

Contact

For questions or issues:

• Dataset repository: https://huggingface.co/datasets/COINjecture/NP-Solutions
• Open a discussion on the dataset page

Changelog

2026-04-16

• Hub CastError / “column names don’t match”: Older JSONL omitted optional fields entirely (serde skip_serializing_if). Different shards then inferred different Arrow column sets. Nodes now serialize every DatasetRecord field on every line (null when None), so new shards align with the full schema. Existing Hub files stay sparse until replaced or batch-normalized.

2026-04-15

• JSONL shape (nodes): mining_attempts is always serialized (use null when unknown); consensus rows set it from header nonce. problem_data / solution_data are coerced to JSON objects before upload so stringified JSON blobs are not emitted.
• Dataset card YAML: Do not add a configs / on_mixed_types block for this repo: it made the Hub builder infer problem_data/solution_data as strings while legacy shards still use nested JSON objects, which triggers DatasetGenerationError / CastError when generating the preview table.

2026-04-10

• explorer_card field: Each JSONL row includes a precomputed multi-line card (UTC time); implemented in huggingface/src/explorer_card.rs.
• Explorer layout: Documented the block-card presentation (block #, type, time, problem/solution prose, solver, BEANS reward, work, asymmetry, quality, Δt, energy) with line-by-line JSON mapping and a Python format_block_card helper.

2025-11-23

• Unified Dataset: Consolidated all problem types (SubsetSum, SAT, TSP, Custom) into a single continuous dataset
• Schema Fix: Fixed u128 bounty serialization (now serialized as string to avoid JSON precision loss)
• Data Provenance: Added institutional-grade data provenance fields (metrics_source, measurement_confidence, data_version)
• Unified Buffer: Changed from per-problem-type buffers to unified buffer that flushes all types together
• Enhanced Metrics: All consensus blocks now include actual block header metrics (high confidence)