nraptisss/telecom-intent-config-sft-10k

Telecom Intent→Config SFT Dataset (10K)

The first open SFT dataset for training LLMs to translate natural language network intents into structured 5G/6G configurations.

This dataset addresses the #1 gap identified in the telecom LLM research landscape: there is no public training dataset for intent-to-policy translation. All existing telecom datasets (TeleQnA, ORANBench-13K, 6G-Bench) are MCQ evaluation benchmarks — not instruction-following format. This dataset fills that gap.

What This Dataset Contains

10,000 samples, each containing a natural language network intent paired with a schema-valid JSON configuration across 4 target layers of the telecom stack:

Target Layer	Schema Source	% of Dataset	Description
CAMARA NetworkSliceBooking	CAMARA Project	35%	High-level slice booking (sliceType, QoS profile, area, duration)
O-RAN A1 Policy	O-RAN WG2 A1 spec	25%	Enforceable RAN policy (QoS objectives, PRB allocation, scheduler weight)
3GPP Intent (TS 28.312)	3GPP TS 28.312 v18	25%	Formal intent model (intentExpectation, targets, contexts)
O1 NRM Config (TS 28.541)	3GPP TS 28.541 NR NRM	15%	RAN config (ManagedElement → GNBDUFunction → NRCellDU → rrmPolicyMemberList)

Dataset Format

Native ChatML/messages format — directly compatible with TRL SFTTrainer, OpenAI fine-tuning, and any chat-template-based training pipeline.

{
  "messages": [
    {"role": "system", "content": "You are a 5G/6G network orchestrator..."},
    {"role": "user", "content": "Provision a URLLC slice for remote surgery in Hospital Campus with 1 ms latency and 99.999% reliability for 4 hours."},
    {"role": "assistant", "content": "{\"sliceType\": \"URLLC\", \"sliceProfile\": {\"latencyTarget\": 1, ...}}"}
  ],
  "target_layer": "camara",
  "slice_type": "URLLC",
  "use_case": "remote surgery",
  "sector": "healthcare",
  ...
}

Slice Type Distribution

Slice Type	Count	%	3GPP SST	Example Use Cases
mMTC	~3,000	30%	SST=3	Smart city sensors, utility metering, environmental monitoring
eMBB	~2,500	25%	SST=1	4K streaming, cloud gaming, AR/VR, holographic telepresence
URLLC	~2,450	25%	SST=2	Remote surgery, industrial robots, autonomous drones
V2X	~1,000	10%	SST=1	Vehicle platooning, collision avoidance, traffic optimization
MPS	~1,000	10%	SST=2	First responder comms, disaster recovery, border surveillance

Sector Coverage

Manufacturing, healthcare, logistics, municipal, transport, media, agriculture, energy, defense, education, environmental, entertainment, security, real-estate, and more — across 20 geographic areas.

Schema Compliance

• 100% JSON validity — every assistant response parses as valid JSON
• 100% schema compliance — every config contains all required fields for its target layer
• 100% unique — no duplicate intent-config pairs
• 10,000 unique NL intents with diverse phrasing (12 templates per layer × randomized parameters)

Target Config Schemas

Layer 1: CAMARA NetworkSliceBooking

{
  "sliceType": "URLLC",
  "sliceProfile": {
    "latencyTarget": 1,
    "reliability": 99.999,
    "throughput": {"downlink": 50, "uplink": 10},
    "maxNumberOfUEs": 100
  },
  "areaOfService": {"region": "Hospital Campus", "cells": ["cell-020", "cell-021"]},
  "duration": {"startTime": "2025-09-14T12:00:00Z", "endTime": "2025-09-14T16:00:00Z", "durationHours": 4}
}

Layer 2: O-RAN A1 Policy

{
  "policyTypeId": "ORAN_QoE_1.0.0",
  "policyId": "uuid",
  "scope": {"sliceId": {"sst": 2, "sd": "000001"}, "cellIdList": ["cell-020", "cell-021"]},
  "statement": {
    "qosObjectives": {"dlThptReq": 50, "ulThptReq": 10, "pktDelayBudget": 1, "pktLossRate": 0.00001, "availabilityTarget": 99.999},
    "resourceAllocation": {"prbQuotaMin": 15, "prbQuotaMax": 50, "schedulerWeight": 9},
    "enforcementAction": "SLICE_PRB_QUOTA"
  }
}

Layer 3: 3GPP TS 28.312 Intent

{
  "intentId": "uuid",
  "intentExpectation": [{
    "expectationId": "uuid",
    "expectationObject": {"objectType": "NetworkSlice", "objectContexts": [{"contextAttribute": "area", "contextCondition": "WITHIN", "contextValueRange": "Hospital Campus"}]},
    "expectationTargets": [
      {"targetName": "DLThptPerSlice", "targetCondition": "IS_GREATER_THAN", "targetValueRange": 50, "targetUnit": "Mbps"},
      {"targetName": "LatencyPerSlice", "targetCondition": "IS_LESS_THAN", "targetValueRange": 1, "targetUnit": "ms"}
    ],
    "expectationContexts": [{"contextAttribute": "serviceType", "contextCondition": "IS_EQUAL_TO", "contextValueRange": "URLLC"}]
  }],
  "intentFulfillmentInfo": {"fulfillmentStatus": "NOT_YET_FULFILLED"}
}

Layer 4: O1 NR NRM Config (TS 28.541)

{
  "ManagedElement": {
    "id": "gNB-001",
    "GNBDUFunction": {
      "id": "DU-001",
      "NRCellDU": {
        "id": "cell-020",
        "operationalState": "ENABLED",
        "administrativeState": "UNLOCKED",
        "rrmPolicyMemberList": [{
          "sNSSAI": {"sst": 2, "sd": "000001"},
          "rrmPolicyMinRatio": 15,
          "rrmPolicyMaxRatio": 50,
          "rrmPolicyDedicatedRatio": 25
        }]
      }
    }
  }
}

Quick Start

from datasets import load_dataset

ds = load_dataset("nraptisss/telecom-intent-config-sft-10k")

# SFT training with TRL
from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
    model="Qwen/Qwen2.5-7B-Instruct",
    train_dataset=ds["train"],
    eval_dataset=ds["test"],
    args=SFTConfig(
        output_dir="./telecom-intent-model",
        num_train_epochs=3,
        per_device_train_batch_size=4,
        learning_rate=2e-5,
        push_to_hub=True,
    ),
)
trainer.train()

# Filter by layer or slice type
urllc_camara = ds["train"].filter(
    lambda x: x["slice_type"] == "URLLC" and x["target_layer"] == "camara"
)

Metadata Columns

Column	Type	Description
id	string	Unique sample ID
messages	list[dict]	ChatML messages (system, user, assistant)
target_layer	string	camara, a1_policy, intent_3gpp, o1_nrm
slice_type	string	eMBB, URLLC, mMTC, V2X, MPS
sst	int	3GPP S-NSSAI Slice/Service Type (1, 2, or 3)
sd	string	3GPP S-NSSAI Slice Differentiator
use_case	string	Specific use case (e.g., "remote surgery")
sector	string	Industry sector (e.g., "healthcare")
region	string	Geographic area name
latency_ms	float	Target latency in milliseconds
reliability_pct	float	Target reliability percentage
dl_throughput_mbps	float	Target downlink throughput in Mbps
ul_throughput_mbps	float	Target uplink throughput in Mbps
max_ues	int	Maximum number of connected devices

Research Context

This dataset was designed based on schemas and methodologies from:

Paper	Contribution to this dataset
ORION (2025)	CAMARA NetworkSliceBooking schema, A1 policy format, intent lifecycle
TelecomGPT (2024)	3-stage domain adaptation pipeline design, telecom task taxonomy
ORANSight 2.0 (2025)	QLoRA instruction-tuning from specs, RANSTRUCT methodology
TSLAM-Mini (2025)	QLoRA training config, chat template design
NEFMind (2025)	Telecom API automation, synthetic data from API specs
3GPP TS 28.312	Intent data model (intentExpectation, targets, contexts)
3GPP TS 28.541	NR NRM YANG model (ManagedElement, NRCellDU, rrmPolicyMemberList)
O-RAN WG2/WG3	A1 policy JSON, E2SM-RC control actions

Limitations

• Synthetic data: Generated from templates with controlled randomization, not from real operator intents. Real-world intents are more ambiguous and underspecified.
• Schema completeness: Covers the core fields of each schema but not every optional field (e.g., CAMARA's notificationCallbackUrl, A1's validity period in some samples).
• No multi-turn: Each sample is a single turn. Real intent lifecycles involve create→assure→modify→terminate dialogues.
• English only: All intents are in English.

Citation

@dataset{telecom_intent_config_sft_10k,
  title={Telecom Intent-Config SFT Dataset (10K)},
  author={nraptisss},
  year={2025},
  url={https://huggingface.co/datasets/nraptisss/telecom-intent-config-sft-10k},
  note={Synthetic dataset for training LLMs on 5G/6G intent-to-config translation}
}

License

Apache 2.0