OTel-Embedding-4B
OTel-Embedding-4B is a telecom embedding model full-parameter fine-tuned on OTel telecommunications data. It is part of the
OTel Family of Models, an open-source initiative to build reference AI resources for the global telecommunications sector.
Across the OTel embedding baselines, OTel fine-tuning improves telecom retrieval quality by +9.6 to +60.2 NDCG@10 percentage points.
Community Use
As of June 23, 2026, the released OTel models had more than 18 million downloads, and the Open Telco AI project had received 157+ pieces of media coverage worldwide.
Model Details
| Attribute | Value |
|---|
| Base model | Qwen/Qwen3-Embedding-4B |
| Parameters | 4B |
| OTel training dataset | OTel-Embedding |
| Dataset fields | anchor, positive, negative_1 through negative_5 |
| Training method | Full-parameter post-training / fine-tuning |
| Language | English |
| OTel release license | Apache 2.0 |
Model Lineage
Qwen/Qwen3-Embedding-4B -> OTel-Embedding full-parameter post-training -> farbodtavakkoli/OTel-Embedding-4B
OTel vs. Base Model
| Metric | Base model | OTel fine-tuned | Delta | Evaluation split |
|---|
| NDCG@10 | 82.5% | 92.2% +/- 0.4 | +9.7 pp | OTel-Embedding held-out 10% |
Standard errors are computed with bootstrap resampling (n=10) over the held-out OTel evaluation partition. NDCG@10 measures whether relevant telecom passages are ranked near the top of the retrieved list.
Evaluation Caveats
- Embedding results measure held-out OTel retrieval partitions.
- Reported standard errors come from bootstrap resampling over the held-out evaluation partitions.
- Aggregate NDCG@10 does not guarantee uniform performance across all telecom subdomains.
- External benchmark transfer, multilingual performance, and per-subdomain performance should be evaluated separately for production settings.
Training Data
The model was trained on telecom-focused data curated by 100+ domain experts. The raw corpus contained roughly 1.1M training points and was filtered to 326,767 higher-confidence examples.
| Source | Contributor |
|---|
| arXiv telecom papers, 3GPP standards, telecom Wikipedia, telecom Common Crawl | Yale University |
| GSMA Permanent Reference Documents, Discover portal | GSMA |
| IETF RFC series | NetoAI |
| Industry whitepapers | Khalifa University |
| O-RAN specifications (working groups 1, 2, 4, 5, 6, 7, 8, 9, 10) | University of Leeds |
| O-RAN documents across working groups | The University of Texas at Dallas |
The OTel datasets release derived QA/retrieval/reranking examples rather than the raw source documents.
Each released dataset includes a dataset card and Croissant metadata with Responsible AI fields for data limitations, biases, sensitive-information considerations, use cases, social impact, synthetic-data status, and provenance.
Representative Training Row
OTel-Embedding rows provide one anchor query, one positive passage, and five hard negatives for telecom retrieval training.
1{
2 "anchor": "During the Measurement ID Coordination (MN initiated) test case, what is the relationship between the U-Plane data and the F1 logs as observed in the Protocol Analyzer and the Test UE or UE emulator?",
3 "positive": "F1 logs recorded in the Protocol Analyzer and the Test UE or UE emulator show that all downlink U-Plane data recorded in the F1 logs is correctly received, and all uplink U-Plane data transmitted by the Test UE or emulated UE is recorded in the F1 logs.",
4 "negative_count": 5
5}
Intended Use
This model is intended to retrieve relevant chunks from telecom specifications, standards, RFCs, whitepapers, and other domain documents. It can be used as the retrieval stage in an end-to-end telecom RAG pipeline, followed by a reranker and a context-grounded LLM.
Training Recipe
| Item | Value |
|---|
| Framework | ScalarLM |
| Optimizer | AdamW, 8-bit |
| Learning-rate schedule | Cosine decay with warmup |
| Weight decay | 0.01 |
| Warmup steps | 100 |
| Random seed | 42 |
| Maximum sequence length | 1500 tokens |
| Precision | BF16 |
| Attention | Flash Attention 2 |
| Distributed training | Fully Sharded Data Parallel |
| Gradient checkpointing | Enabled |
| Epochs | 3 for LLM/embedding models; 2 for rerankers |
| Compute | AMD MI300X/MI325X/MI355X and NVIDIA A100/H100 GPUs |
Usage
1from sentence_transformers import SentenceTransformer
2
3model = SentenceTransformer("farbodtavakkoli/OTel-Embedding-4B")
4
5sentences = [
6 "What is the F1 interface in O-RAN?",
7 "The F1 interface connects the O-RAN Distributed Unit (O-DU) to the O-RAN Central Unit (O-CU).",
8]
9
10embeddings = model.encode(sentences, normalize_embeddings=True)
11print(embeddings.shape)
Limitations and Responsible Use
- OTel models are domain-specific to telecommunications and should not be treated as general-purpose models.
- The current release is English-only and primarily text-centric.
- The reported OTel performance results use held-out OTel evaluation partitions and should not be interpreted as results from a fully independent external benchmark suite.
- Aggregate scores can hide subdomain variation; collaborator stress tests suggest O-RAN retrieval is comparatively strong, while academic-paper and GSMA PRD examples need further curation.
- Generated telecom content should be verified before operational, customer-facing, regulatory, safety, or network-configuration use.
- Users must comply with both the OTel release license and the upstream base-model license or terms.
Related Models
Project Resources
Citation
1@misc{otel_models_2026,
2 title = {OTel: Open Telco AI Datasets, Benchmarks, and Models},
3 author = {Tavakkoli, Farbod and others},
4 year = {2026},
5 note = {Open Telco (OTel) model release},
6 url = {https://huggingface.co/farbodtavakkoli}
7}
Contact
