README.md
| 1 | --- |
| 2 | license: apache-2.0 |
| 3 | language: |
| 4 | - en |
| 5 | base_model: |
| 6 | - google/gemma-4-E4B-it |
| 7 | datasets: |
| 8 | - farbodtavakkoli/OTel-LLM |
| 9 | tags: |
| 10 | - telecom |
| 11 | - telecommunications |
| 12 | - gsma |
| 13 | - rag |
| 14 | - full-parameter-fine-tuning |
| 15 | - fine-tuned |
| 16 | pipeline_tag: text-generation |
| 17 | --- |
| 18 | |
| 19 | # OTel-LLM-E4B-IT |
| 20 | |
| 21 | **OTel-LLM-E4B-IT** is a context-grounded telecom language model full-parameter fine-tuned on OTel telecommunications data. It is part of the [OTel Family of Models](https://huggingface.co/collections/farbodtavakkoli/otel-llm), an open-source initiative to build reference AI resources for the global telecommunications sector. |
| 22 | |
| 23 | Across the core OTel LLM baselines, OTel fine-tuning improves context-grounded correctness over the base checkpoints by +3.7 to +10.0 percentage points. |
| 24 | |
| 25 | ## Community Use |
| 26 | |
| 27 | 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. |
| 28 | |
| 29 | ## Model Details |
| 30 | |
| 31 | | Attribute | Value | |
| 32 | |---|---| |
| 33 | | Base model | [google/gemma-4-E4B-it](https://huggingface.co/google/gemma-4-E4B-it) | |
| 34 | | Parameters | 4.5B | |
| 35 | | OTel training dataset | [OTel-LLM](https://huggingface.co/datasets/farbodtavakkoli/OTel-LLM) | |
| 36 | | Dataset fields | `prompt`, `completion`, `abstention`, chunk-count metadata, token-count metadata | |
| 37 | | Training method | Full-parameter post-training / fine-tuning | |
| 38 | | Language | English | |
| 39 | | OTel release license | Apache 2.0 | |
| 40 | |
| 41 | ## Model Lineage |
| 42 | |
| 43 | `google/gemma-4-E4B-it` -> `OTel-LLM` full-parameter post-training -> `farbodtavakkoli/OTel-LLM-E4B-IT` |
| 44 | |
| 45 | ## OTel vs. Base Model |
| 46 | |
| 47 | | Metric | Base model | OTel fine-tuned | Delta | Evaluation split | |
| 48 | |---|---:|---:|---:|---| |
| 49 | | LLM-as-judge correctness | 82.4% | 91.7% +/- 0.4 | +9.3 pp | OTel-LLM held-out 10% | |
| 50 | |
| 51 | Standard errors are computed with bootstrap resampling (`n=10`) over the held-out OTel evaluation partition. LLM correctness is judged by GPT-4o mini using the retrieved context and reference answer. |
| 52 | |
| 53 | ## Evaluation Caveats |
| 54 | |
| 55 | - LLM results measure context-grounded answer generation from retrieved context, not unrestricted context-free telecom QA. |
| 56 | - Reported standard errors come from bootstrap resampling over the held-out evaluation partitions. |
| 57 | - Answer quality depends on the retriever, reranker, context window, and prompt policy around the model. |
| 58 | - External benchmark transfer, multilingual performance, and per-subdomain performance should be evaluated separately for production settings. |
| 59 | |
| 60 | |
| 61 | ## Training Data |
| 62 | |
| 63 | 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. |
| 64 | |
| 65 | | Source | Contributor | |
| 66 | |---|---| |
| 67 | | arXiv telecom papers, 3GPP standards, telecom Wikipedia, telecom Common Crawl | Yale University | |
| 68 | | GSMA Permanent Reference Documents, Discover portal | GSMA | |
| 69 | | IETF RFC series | NetoAI | |
| 70 | | Industry whitepapers | Khalifa University | |
| 71 | | O-RAN specifications (working groups 1, 2, 4, 5, 6, 7, 8, 9, 10) | University of Leeds | |
| 72 | | O-RAN documents across working groups | The University of Texas at Dallas | |
| 73 | |
| 74 | Released datasets: [OTel-LLM](https://huggingface.co/datasets/farbodtavakkoli/OTel-LLM), [OTel-Embedding](https://huggingface.co/datasets/farbodtavakkoli/OTel-Embedding), [OTel-Reranker](https://huggingface.co/datasets/farbodtavakkoli/OTel-Reranker), and [OTel-Safety](https://huggingface.co/datasets/farbodtavakkoli/OTel-Safety). |
| 75 | |
| 76 | The OTel datasets release derived QA/retrieval/reranking examples rather than the raw source documents. |
| 77 | |
| 78 | 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. |
| 79 | |
| 80 | ## Representative Training Row |
| 81 | |
| 82 | `OTel-LLM` rows pair a context-grounded telecom RAG prompt with a reference completion. |
| 83 | |
| 84 | ```json |
| 85 | { |
| 86 | "anchor": "How can a cell be considered to be operating in MBSFN mode for 3.84/7.68 Mcps TDD?", |
| 87 | "completion": "A cell shall be considered to be operating in MBSFN mode when individual scrambling codes are assigned to all timeslots via the IE \"TDD MBSFN Information\".", |
| 88 | "abstention": false, |
| 89 | "n_positive_chunks": 1, |
| 90 | "n_negative_chunks": 4 |
| 91 | } |
| 92 | ``` |
| 93 | |
| 94 | ## Intended Use |
| 95 | |
| 96 | This model is intended for context-grounded telecom answer generation in Retrieval-Augmented Generation (RAG) pipelines. It should receive retrieved telecom context and generate an answer grounded in that context. |
| 97 | |
| 98 | The model is not optimized for unrestricted context-free question answering. For questions where the retrieved context is missing or insufficient, use an abstention-aware prompt or one of the dedicated `-Safety` variants. |
| 99 | |
| 100 | ## Training Recipe |
| 101 | |
| 102 | | Item | Value | |
| 103 | |---|---| |
| 104 | | Framework | ScalarLM | |
| 105 | | Optimizer | AdamW, 8-bit | |
| 106 | | Learning-rate schedule | Cosine decay with warmup | |
| 107 | | Weight decay | 0.01 | |
| 108 | | Warmup steps | 100 | |
| 109 | | Random seed | 42 | |
| 110 | | Maximum sequence length | 1500 tokens | |
| 111 | | Precision | BF16 | |
| 112 | | Attention | Flash Attention 2 | |
| 113 | | Distributed training | Fully Sharded Data Parallel | |
| 114 | | Gradient checkpointing | Enabled | |
| 115 | | Epochs | 3 for LLM/embedding models; 2 for rerankers | |
| 116 | | Compute | AMD MI300X/MI325X/MI355X and NVIDIA A100/H100 GPUs | |
| 117 | |
| 118 | ## Usage |
| 119 | |
| 120 | ```python |
| 121 | from transformers import AutoModelForCausalLM, AutoTokenizer |
| 122 | import torch |
| 123 | |
| 124 | model_name = "farbodtavakkoli/OTel-LLM-E4B-IT" |
| 125 | tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) |
| 126 | model = AutoModelForCausalLM.from_pretrained( |
| 127 | model_name, |
| 128 | torch_dtype=torch.bfloat16, |
| 129 | device_map="auto", |
| 130 | trust_remote_code=True, |
| 131 | ) |
| 132 | |
| 133 | prompt = """You are a precise telecom assistant in a RAG pipeline. |
| 134 | Use only the retrieved context to answer. |
| 135 | |
| 136 | User Question |
| 137 | What is the purpose of the F1 interface in O-RAN? |
| 138 | |
| 139 | Retrieved Contexts |
| 140 | CONTEXT 1 |
| 141 | The F1 interface connects the O-RAN Distributed Unit (O-DU) to the O-RAN Central Unit (O-CU). |
| 142 | |
| 143 | Answer:""" |
| 144 | |
| 145 | inputs = tokenizer(prompt, return_tensors="pt").to(model.device) |
| 146 | outputs = model.generate(**inputs, max_new_tokens=256) |
| 147 | print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
| 148 | ``` |
| 149 | |
| 150 | ## Limitations and Responsible Use |
| 151 | |
| 152 | - OTel models are domain-specific to telecommunications and should not be treated as general-purpose models. |
| 153 | - The current release is English-only and primarily text-centric. |
| 154 | - 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. |
| 155 | - 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. |
| 156 | - Generated telecom content should be verified before operational, customer-facing, regulatory, safety, or network-configuration use. |
| 157 | - Users must comply with both the OTel release license and the upstream base-model license or terms. |
| 158 | - For unrestricted telecom QA without retrieved context, use a separately evaluated context-free QnA model rather than assuming this RAG-oriented checkpoint will behave optimally. |
| 159 | |
| 160 | ## Related Models |
| 161 | |
| 162 | - [OTel LLM Collection](https://huggingface.co/collections/farbodtavakkoli/otel-llm) |
| 163 | - [OTel Embedding Collection](https://huggingface.co/collections/farbodtavakkoli/otel-embedding) |
| 164 | - [OTel Reranker Collection](https://huggingface.co/collections/farbodtavakkoli/otel-reranker) |
| 165 | |
| 166 | ## Project Resources |
| 167 | |
| 168 | - Project page: https://huggingface.co/farbodtavakkoli |
| 169 | - Code: https://github.com/farbodtavakkoli/OTel |
| 170 | - Media coverage list: https://github.com/farbodtavakkoli/OTel/blob/main/docs/media_coverage.md |
| 171 | |
| 172 | ## Citation |
| 173 | |
| 174 | ```bibtex |
| 175 | @misc{otel_models_2026, |
| 176 | title = {OTel: Open Telco AI Datasets, Benchmarks, and Models}, |
| 177 | author = {Tavakkoli, Farbod and others}, |
| 178 | year = {2026}, |
| 179 | note = {Open Telco (OTel) model release}, |
| 180 | url = {https://huggingface.co/farbodtavakkoli} |
| 181 | } |
| 182 | ``` |
| 183 | |
| 184 | ## Contact |
| 185 | |
| 186 | For technical questions, contact farbod.tavakkoli@att.com or farbodtavakoli@gmail.com. |
| 187 | |