saidurga001301/agent-primitives-pilot-1k

Agent Primitives Pilot: High-Entropy Extraction Dataset (305k Rows)

A high-fidelity agentic extraction dataset specifically curated and aligned for small, ultra-fast language and hybrid models (e.g., Liquid AI LFM2-1.2B-Extract and LFM3-1.7B).

All synthetic degeneration loops, streaming repetitions, and padded artifacts have been permanently pruned. Every target strictly satisfies the $\le 32$ token ceiling to force models into concise, hallucination-free primitive extraction.

---

1. Composition Summary (305,063 Total Clean Rows)

Primitive	Samples	Share	Description & Benchmark Source	Target Ceiling
INT	99,162	32.51%	Deterministic integer answers (MetaMathQA)	$\le 12$ tokens
SPAN	86,580	28.38%	Verbatim factual document answer spans (SQuAD v2)	$\le 32$ tokens
STRUCT	59,270	19.43%	Compact JSON key-value extraction (LiquidAI/structured-extraction)	$\le 32$ tokens
CODE	28,769	9.43%	Linux commands & tool-calling arguments (Romit2004/LinuxCommands, Glaive)	$\le 32$ tokens
STR	9,575	3.14%	Scientific and technical string extraction (SciQ, IETF RFCs)	$\le 20$ tokens
NOUL	8,553	2.80%	Explicit non-hallucination abstentions ("null", BoolQ & missing schema fields)	1 token
CHOICE	7,214	2.36%	Fast decision routing & tool name selection (Fastino fast-decisions)	$\le 6$ tokens
REPLAY	5,940	1.95%	Exact reproduction & verbatim state recovery (Databricks Dolly)	$\le 32$ tokens

---

2. Distributed Cluster Execution (distlab Orchestration)

This dataset was processed and verified using the distlab distributed cluster engine across 32 idle compute nodes (384 CPU cores):

• Master Node: 10.5.16.86 (orchestration, slicing, parquet encoding, and upload).
• Worker Execution: 32 nodes executed pure standard-library workers strictly inside ~/.work/ without touching host directories.
• Batch Dispatch: 32 jobs dispatched concurrently via distlab submit --batch (Batch ID: b_1031_185710_e3da).
• Runtime: Completed in 4.8 seconds across 32 nodes (32/32 successful, 0 failed).
• Target Verification: 0 target token violations (>32 tokens = 0.00%).

---

3. Quickstart & Usage

from datasets import load_dataset

# Load compact high-entropy shards directly:
ds = load_dataset("saidurga001301/agent-primitives-pilot-1k", data_dir="data_compact", split="train")

print(f"Total Rows: {len(ds):,}")
print("Sample:", ds[0])