
Base model	`AliMaatouk/LLama-3-8B-Tele` (Llama-3-8B pretrained on telecom corpora)
Adapter type	QLoRA (4-bit NF4 quantized, r=64, α=128, dropout=0.05)
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Trainable params	167M / 8.2B (2.05%)
Training data	18,779 synthetic instruction–response pairs across 8 TMF-aligned categories
Training config	3 epochs · effective batch 16 · seq 4096 · cosine LR (peak 2e-4) · bf16 · gradient checkpointing
Hardware	NVIDIA A100 SXM4 80GB · ~8.3 hours wall time
Best eval loss	0.8438 (epoch 2.56) — `load_best_model_at_end=True`

Intended use

Domain-specialised completions and code generation for:

TMF Open API payload generation (TMF620–TMF700 suite)
eTOM process decomposition (Fulfillment / Assurance / Billing end-to-end flows)
SID entity relationship reasoning (ProductOffering → Service → Resource hierarchies, Party/Role patterns, characteristic specifications)
Inventory reconciliation (discovery–inventory mismatches, phantom/orphan resources)
OSS/BSS architecture decisions (ODA components, build-vs-buy, MANO choices)
Fault-to-inventory correlation (service impact from topology traversal)
TMF spec Q&A (technical knowledge retrieval)
Integration code (TMF-compliant Python clients)

How to use

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "AliMaatouk/LLama-3-8B-Tele"
adapter = "Tapask/telecom-oss-8b"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, adapter)
model.eval()

prompt = """Below is an instruction that describes a task related to telecom OSS/BSS systems, TMF Frameworx, or network operations. Write a response that appropriately completes the request.

### Instruction:
Generate a TMF641 service order payload for a 5G network slice with URLLC characteristics targeting an enterprise IoT customer.

### Response:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=1024, temperature=0.3, do_sample=True)
print(tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Uses the Alpaca prompt template the model was trained with. Keep the ### Instruction: / ### Response: markers exactly.

Training data

18,779 instruction–response pairs were generated programmatically via the Claude API, Kimi K2.5 on Ollama Cloud, and GLM-5 on Ollama Cloud, prompted with 8 category-specific TMF expert personas (system prompts + 4–5 batch variants each). Distribution:

#	Category	Pairs	Primary model
1	TMF Open API Payloads	2,962	GLM-5
2	eTOM Process Decomposition	1,967	GLM-5
3	SID Entity Reasoning	1,963	Kimi K2.5
4	Inventory Reconciliation	2,962	Kimi K2.5
5	OSS/BSS Architecture	1,893	Kimi K2.5
6	Fault-to-Inventory Correlation	1,929	GLM-5
7	TMF Spec Q&A	2,875	Kimi K2.5 (after GLM-5 hit 54% dedup rate)
8	TMF Integration Code Generation	2,228	GLM-5

Splits (seed 42): 16,901 train / 939 val / 939 test.

Quality passes applied:

MD5-hash deduplication on instruction field
Category-aware soft validators (TMF API reference presence, SID entity coverage, eTOM term coverage, JSON validity for payload categories)
Refusal-pattern scrubbing (I cannot, As an AI, etc. removed)
Type coercion for 297 pairs where source models emitted output as nested JSON objects instead of JSON strings

Evaluation loss trajectory

Epoch	Eval loss
2.27	0.8545
2.37	0.8440
2.46	0.8447
2.56	0.8438 ← best, used for final adapter
2.65	0.8479
2.75	0.8478

Loss plateaued and began ticking up after epoch 2.56 — classic mild overfitting signal. load_best_model_at_end=True ensured the final adapter corresponds to the epoch 2.56 region.

Limitations

Synthetic data provenance — training pairs were generated by LLMs (Claude, Kimi K2.5, GLM-5) prompted with TMF expert personas. Content is stylistically consistent with TMF specs but not validated line-by-line against official TMF Open API documents. Treat outputs as starting points, not canonical.
Inner-JSON flaws — ~15% of category-1 pairs had minor inner-JSON issues (unescaped quotes inside payload strings). Not filtered out for training.
Category 8 undertrained — TMF Code Generation category ended at 74% of its 3,000-pair target due to narrow topic space and dedup loss. Code-generation quality is the weakest axis.
Domain scope — the model is narrow. General-purpose conversation, math, code outside TMF integration, etc. will be no better (and often worse) than the base.
Standards currency — trained against TMF Open API versions current as of the prompt design (~v4/v5 dominant). May cite outdated endpoint paths for newer TMF releases.

Intended use — restrictions

Follows the Llama 3 Community License and Acceptable Use Policy. Additionally, this adapter is intended for:

Domain research, prototyping, and educational use
Assistant-style answers to TMF/OSS/BSS engineering questions
Starter payload generation (to be reviewed before use in production)

Not suitable for:

Generating production systems config without human review
Compliance-sensitive deployments (TMF spec accuracy is not guaranteed)
Any of the prohibited uses in the Llama 3 AUP

License

Adapter weights: inherit Llama 3 Community License from the base model
Attribution required: see top of this card
Note that Llama 3 license restricts some commercial uses (700M+ MAU clause) and prohibited use cases — consult the license before redistribution

Acknowledgements

Meta AI — Llama 3 base model
Ali Maatouk — telecom-pretrained continuation AliMaatouk/LLama-3-8B-Tele
Anthropic, Moonshot AI, Zhipu AI — Claude, Kimi K2.5, GLM-5 (used to generate synthetic training data)
TMForum — the eTOM, SID, and Open API standards this model targets

Citation

@misc{tapask_telecom_oss_8b_2026,
  title        = {Telecom OSS/BSS Domain LLM (LoRA Adapter for LLama-3-8B-Tele)},
  author       = {Tapas},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Tapask/telecom-oss-8b}},
}