| --- |
| license: apache-2.0 |
| base_model: unsloth/Qwen2.5-7B-Instruct-bnb-4bit |
| tags: |
| - telecom |
| - ericsson |
| - huawei |
| - nokia |
| - zte |
| - 5g |
| - 5g-nr |
| - lte |
| - ran |
| - 5gc |
| - oran |
| - fine-tuned |
| - qwen2.5 |
| - unsloth |
| language: |
| - en |
| library_name: transformers |
| pipeline_tag: text-generation |
| model-index: |
| - name: telecom-intelligence-model-v7-merged |
| results: [] |
| --- |
| |
| # Telecom Intelligence — Qwen2.5-7B v7 |
|
|
| A domain-fine-tuned LLM for telecom network operations, built on [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) using QLoRA (4-bit) + SFT via [Unsloth](https://github.com/unslothai/unsloth). |
|
|
| This is the **merged, standalone model** — no adapter loading required. Compatible with vLLM, Transformers, and HuggingFace Inference Endpoints. |
|
|
| **v7 adds 200 new training examples** (805 total) covering 5G NR deep internals, multi-vendor PM counter diagnosis, and advanced 5GC NF fault chains — the largest training set in this series. |
|
|
| --- |
|
|
| ## What it does |
|
|
| The model reasons step-by-step over telecom operational data to: |
|
|
| - **Root cause analysis** — diagnose KPI degradations from PM counter data across Ericsson, Huawei, Nokia, and ZTE RAN/Core nodes |
| - **5G NR deep knowledge** — SSB beam management (P1/P2/P3/BFR), NR numerology (μ=0–4, SCS, slot duration, PRBs per bandwidth), CORESET/PDCCH blind decoding, BWP switching, PUCCH formats 0–4 (HARQ-ACK/SR/CSI), SIB1 contents, F1/E1 interface (CU/DU split), RRC_INACTIVE state (I-RNTI, RNA, Resume), SDAP layer (QoS flow→DRB mapping) |
| - **LTE RAN** — RRC/ERAB/HO KPI chains with correct Ericsson `pmRrcConnEstabSucc`/`pmErabEstabSuccInit`/`pmHoExeSuccLteIntraF` formulas; inter-frequency HO (A2/A3/A4/A5 events, measurement gaps); LTE TA/TAU; eICIC/ABS (HetNet, CRE, FeICIC); S1 release causes; RSRP→SINR→CQI→MCS link adaptation chain with OLLA |
| - **5G Core NF attribution** — names the exact failing NF (AMF, SMF, UPF, PCF, AUSF, UDM) and interface (N4/PFCP, N8, N11, N7, N10, NGAP) rather than vague "core network" answers; AMF overload scenarios, SMF/PFCP session failure RCA, UPF pod crash diagnosis |
| - **PRB utilisation** — correct formula for all LTE bandwidths and 5G NR; differentiates congestion vs RF root cause |
| - **SON Energy Saving** — binary ACTIVATE / DO NOT ACTIVATE decisions with threshold reasoning (PRB, UE count, neighbour overlap, NOC approval) |
| - **Multi-vendor counter normalisation** — Ericsson `pm*`, Huawei `L.*` / `VS.5G.*`, Nokia `RRC_CONN_*`, ZTE LTE PM naming conventions; maps all to equivalent KPI formulas |
| - **Huawei MML** — canonical command verbs: `BLK`/`UBL`, `LST`, `DSP`, `MOD`, `RST`, `ACT`, `DEA`, `ADD`, `RMV`; includes `RST NRDUCELL` (5G DU cell reset procedure) |
| - **Ericsson AMOS CLI** — `get`/`set`/`la`/`st` commands; targets correct MO classes (`EUtranCellFDD`, `NRCellDU`, `AntennaUnitGroup`, `RetSubUnit`); uses `administrativeState` attribute |
| - **EN-DC / NR-DC / NSA vs SA** — architecture differences, X2/Xn procedures, EN-DC setup failure diagnosis, UE capability procedure |
| - **PDCP lossless handover** — SN Status Transfer, PDCP SDU forwarding, in-sequence delivery guarantee |
| - **A3 event HO parameters** — a3-Offset, hysteresis, TTT, MRO; UL power control (P0/alpha/TPC); HSDPA vs LTE L2 scheduling differences |
| - **DL spectral efficiency** — derived from `pmPdcpVolDlDrb` + `pmPrbUsedDlSum` with correct 18 MHz used BW for 20 MHz LTE |
| - **ENM vs ENIQ analytics** — when to use each, ENIQ SQL example, KPI cooking vs ad-hoc troubleshooting |
| - **Nokia NetAct 5G** — RACT CLI, BTS Manager CLI, MO hierarchy (GNBDU/GNBCUCP/GNBCUUP), PM counter access |
| - **O-RAN, VoLTE/VoNR, IMS, NTN, cloud-native NF** — coverage inherited from v5/v6 |
| |
| --- |
| |
| ## Training details |
| |
| | Parameter | Value | |
| |-----------|-------| |
| | Base model | `unsloth/Qwen2.5-7B-Instruct-bnb-4bit` | |
| | Training examples | 805 | |
| | Training method | QLoRA (4-bit) + SFT via Unsloth + TRL SFTTrainer | |
| | LoRA rank | r=16, alpha=32 | |
| | LoRA target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj | |
| | Max sequence length | 2048 | |
| | Epochs | 3 | |
| | Merge type | 16-bit merged (standalone, no adapter required) | |
|
|
| --- |
|
|
| ## Training data coverage |
|
|
| New in v7 (batches 47–52, ~110 examples): |
|
|
| - LTE TA/TAU, S6a vs N8 interface distinction |
| - Ericsson AMOS commands for LTE and 5G NR |
| - RSRP vs RSRQ — measurement purpose and thresholds |
| - `pmRrcConnEstabFail` by cause code |
| - LTE SCell/CA activation, UE power headroom (PHR) |
| - Huawei iMaster NCE vs U2020 differences |
| - 5G NSA vs SA architecture, EN-DC procedure + counters |
| - AMF overload scenarios and handling |
| - Huawei `VS.5G.*` PM counter families |
| - LTE S1 UE Context Release causes |
| - ZTE LTE PM counter naming convention |
| - 5G network slicing / S-NSSAI |
| - LTE inactivity timer + DRX interaction |
| - LTE X2 interface procedures |
| - LTE UL power control (P0, alpha, TPC) |
| - PDCP lossless HO (SN Status Transfer + forwarding) |
| - A3 event parameters (a3-Offset, hysteresis, TTT, MRO) |
| - eNB vs gNB CU/DU split (F1/E1 interfaces) |
| - WCDMA HSDPA vs LTE L2/scheduling architecture |
| - Nokia GNBCUCP HO counter hierarchy and dashboard |
| - 5G NR PRB high + low throughput diagnosis workflow |
| - NR-DC vs EN-DC differences |
| - 5G NR physical channels and signals (PSS/SSS/PBCH/PDCCH/PDSCH/PUCCH/PRACH/SRS) |
| - LTE E2E session — power-on to first packet (22 steps) |
| - Ericsson `pmErabEstabSuccInitMmeTrigger` (MME-triggered E-RAB) |
| - 5G NR CORESET + PDCCH blind decoding |
| - 5G BWP concept and switching |
| - SSB beam management (P1/P2/P3, BFR) |
| - NGAP message set (NG Setup, InitialUEMessage, PDU session, HO, Paging, Reset) |
| - LTE UE category table (Cat-1 to Cat-NB1, peak rates) |
| - WCDMA R99 AMR vs VoLTE/VoNR evolution |
| - 5G UPF failure scenarios |
| - 5G NR numerology (μ=0–4, SCS implications) |
| - Huawei SmartPower (Symbol/Channel/Cell Shutdown, energy saving counters) |
| - LTE inter-freq HO (A2/A3/A4/A5/A1 events, measurement gaps) |
| - GTP-U protocol (TEID, protocol stack, LTE vs 5G N3 with QFI) |
| - WCDMA soft HO vs LTE hard HO (fundamental CDMA vs OFDMA reason) |
| - RRC drop diagnosis (congestion/PRACH/MME) |
| - SMF role + N4/N7/N10/N11/N40 interfaces |
| - NR RACH 4-step vs 2-step (Type 1 vs Type 2, SSB association) |
| - SDAP layer (QoS flow→DRB mapping, Reflective QoS, CU-UP) |
| - SRS and massive MIMO UL (reciprocity-based beamforming, MU-MIMO scheduling) |
| - Ericsson top 5 5G alarms + actions |
| - 3GPP Rel-17 key features (RedCap, NTN, Sidelink, Positioning, MBS) |
| - LTE DL spectral efficiency calculation from PM counters |
| - 5G NR PUCCH (formats 0–4, UCI: HARQ/SR/CSI) |
| - Ericsson `pmHoPrepSuccLteIntraF` + full HO KPI chain |
| - Ericsson ENIQ vs ENM PM — analytics use-case comparison |
| - Nokia NetAct 5G management (RACT CLI, MO hierarchy, PM counters) |
| - 5G RRC_INACTIVE vs RRC_CONNECTED (resume, I-RNTI, small data) |
| - Ericsson paging SR% (pmPagingAtt/Succ, DRX, TAU) |
| - 5G NR SIB1 contents (cellAccessRelatedInfo, RACH config, TDD offset) |
| - LTE eICIC/ABS (HetNet, ABS bitmap, FeICIC, CRE) |
| - LTE UE capability procedure (SupportedBandCombinationList) |
| - VS.5G.Cell.HO.Fail.RLF — antenna replacement root cause diagnosis |
| - L.HO.Succ.Inter.eNB vs Intra.eNB + X2 vs S1 HO breakdown |
| - 5G F1 interface (F1AP message set, F1-U GTP-U) |
| - Huawei MML `RST NRDUCELL` (cell reset procedure + precautions) |
| - RSRP→SINR→CQI→MCS link adaptation chain (CQI table, OLLA) |
|
|
| --- |
|
|
| ## Quick start |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| import torch |
| |
| model_id = "mindfossil/telecom-intelligence-model-v7-merged" |
| |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, |
| torch_dtype=torch.float16, |
| device_map="auto", |
| ) |
| |
| SYSTEM_PROMPT = ( |
| "You are a telecom network intelligence assistant. " |
| "You analyse probe data, RAN PM counters, Core PM metrics, and transport layer KPIs " |
| "to detect anomalies, diagnose faults, perform root cause analysis, translate natural " |
| "language to queries, and generate AMOS or MML CLI commands. " |
| "Always reason step by step: identify the vendor and counter naming convention, " |
| "compute all KPIs explicitly showing the arithmetic, compare against known thresholds, " |
| "then state the root cause and recommended action." |
| ) |
| |
| def ask(question: str) -> str: |
| messages = [ |
| {"role": "system", "content": SYSTEM_PROMPT}, |
| {"role": "user", "content": question}, |
| ] |
| inputs = tokenizer.apply_chat_template( |
| messages, tokenize=True, add_generation_prompt=True, return_tensors="pt" |
| ).to(model.device) |
| outputs = model.generate( |
| inputs, max_new_tokens=1024, temperature=0.1, do_sample=True |
| ) |
| return tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True) |
| |
| # Example |
| print(ask( |
| "Ericsson cell ENB010-CELL01: pmRrcConnEstabAtt=948, pmRrcConnEstabSucc=921, " |
| "pmErabEstabAttInit=921, pmErabEstabSuccInit=902, pmHoExeAttLteIntraF=421, " |
| "pmHoExeSuccLteIntraF=412. Analyse for anomalies." |
| )) |
| ``` |
|
|
| --- |
|
|
| ## HuggingFace Inference Endpoint (recommended for production) |
|
|
| Deploy as a dedicated endpoint for low-latency inference without managing GPU infrastructure: |
|
|
| ```python |
| import requests, json, os |
| |
| ENDPOINT_URL = "https://<your-endpoint>.aws.endpoints.huggingface.cloud" |
| HF_TOKEN = os.environ["HF_TOKEN"] |
| |
| def ask_endpoint(question: str, system: str = None) -> str: |
| system = system or ( |
| "You are a telecom network intelligence assistant. Reason step by step, " |
| "derive all KPIs showing arithmetic, identify the vendor and counter naming, " |
| "compare against thresholds, then state the root cause and recommended action." |
| ) |
| payload = { |
| "model": "mindfossil/telecom-intelligence-model-v7-merged", |
| "messages": [ |
| {"role": "system", "content": system}, |
| {"role": "user", "content": question}, |
| ], |
| "max_tokens": 1500, |
| "temperature": 0.1, |
| } |
| r = requests.post( |
| f"{ENDPOINT_URL}/v1/chat/completions", |
| headers={"Authorization": f"Bearer {HF_TOKEN}", "Content-Type": "application/json"}, |
| json=payload, timeout=180, |
| ) |
| r.raise_for_status() |
| return r.json()["choices"][0]["message"]["content"] |
| ``` |
|
|
| --- |
|
|
| ## vLLM (self-hosted) |
|
|
| ```bash |
| pip install vllm |
| vllm serve mindfossil/telecom-intelligence-model-v7-merged \ |
| --dtype float16 \ |
| --max-model-len 4096 \ |
| --gpu-memory-utilization 0.90 |
| ``` |
|
|
| --- |
|
|
| ## Example prompts |
|
|
| **PM counter anomaly detection:** |
| ``` |
| TASK: Anomaly Detection |
| VENDOR: Ericsson | Cell: ENB042-CELL07 | Granularity: 15 min |
| |
| pmRrcConnEstabAtt=941, pmRrcConnEstabSucc=908 |
| pmErabEstabAttInit=941, pmErabEstabSuccInit=824 |
| pmHoExeAttLteIntraF=378, pmHoExeSuccLteIntraF=288 |
| |
| Analyse for anomalies. Derive all KPIs showing your arithmetic, then state the verdict. |
| ``` |
|
|
| **5G NR beam failure diagnosis:** |
| ``` |
| After an antenna upgrade at a 5G NR cell, UEs near the cell edge show intermittent |
| disconnections. Logs show frequent BFR (Beam Failure Recovery) events. |
| Explain the SSB beam management procedure (P1/P2/P3), what triggers BFR, |
| and which PM counters to check. |
| ``` |
|
|
| **5GC NF root cause:** |
| ``` |
| SMF-PROD-03: PDU_Session_Estab_SR=58.3% (baseline 99.1%), N4_HeartbeatTimeout_Rate=22.1% |
| (baseline 0%), N11_SR=99.8%, N7_SR=99.6%, SMF_CPU_Util=41%. |
| Diagnose the root cause. Which NF is failing and on which interface? |
| ``` |
|
|
| **Ericsson AMOS CLI:** |
| ``` |
| TASK: AMOS CLI command generation |
| VENDOR: Ericsson | Node: GNBDU-SITE-05 |
| Generate AMOS commands to: |
| 1. Lock all NRCellDU cells (administrativeState) |
| 2. Retrieve pmNrRrcConnEstabSucc and pmNrRrcConnEstabAtt for all cells |
| 3. Unlock NRCellDU=Cell-1 |
| ``` |
|
|
| --- |
|
|
| ## Limitations |
|
|
| - Output quality depends on how explicitly vendor, counter names, and task type are stated in the prompt. The structured prompt format shown above consistently outperforms free-form questions. |
| - The model was not trained on proprietary network configurations or live traffic data. It reasons from 3GPP specifications and publicly available vendor documentation. |
| - Computed KPI values are arithmetic derivations from counter inputs provided in the prompt — the model does not connect to live network systems. |
| - Recommended for augmenting, not replacing, experienced RF/Core network engineers. |
|
|
| --- |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{telecom-intelligence-v7, |
| author = {mindfossil}, |
| title = {Telecom Intelligence Model v7 — Qwen2.5-7B Fine-Tuned for Network Operations}, |
| year = {2026}, |
| publisher = {HuggingFace}, |
| url = {https://huggingface.co/mindfossil/telecom-intelligence-model-v7-merged} |
| } |
| ``` |
|
|
