---
license: apache-2.0
base_model:
- Qwen/Qwen3-8B
datasets:
- nraptisss/TMF921-intent-to-config-research-sota
library_name: peft
tags:
- peft
- qlora
- trl
- sft
- qwen3
- telecom
- tmf921
- 3gpp
- etsi-zsm
- o-ran
- camara
- network-slicing
- intent-based-networking
- json-generation
- research-baseline
pipeline_tag: text-generation
---

# Qwen3-8B TMF921 Intent-to-Configuration QLoRA Adapter

This is the primary stage-1 QLoRA adapter for multi-standard telecom intent-to-configuration translation.

Base model:

- [`Qwen/Qwen3-8B`](https://huggingface.co/Qwen/Qwen3-8B)

Training dataset:

- [`nraptisss/TMF921-intent-to-config-research-sota`](https://huggingface.co/datasets/nraptisss/TMF921-intent-to-config-research-sota)

Training/evaluation repository:

- [`nraptisss/tmf921-intent-training`](https://huggingface.co/nraptisss/tmf921-intent-training)

## Intended task

The model translates natural-language 5G/6G telecom/network-slicing intents into structured JSON configuration-like outputs across target families including:

- TMF921 intent objects,
- 3GPP intent-style objects,
- ETSI ZSM intent-style objects,
- CAMARA network-slice booking-style objects,
- O-RAN A1 policy-style objects,
- O1 NRM-style objects,
- TMF921 lifecycle operations,
- adversarial/rejection responses.

This is a **research baseline**, not a production-certified network-management system.

## Training recipe

| Item | Value |
|---|---|
| Base model | `Qwen/Qwen3-8B` |
| Method | QLoRA SFT |
| Quantization | 4-bit NF4 + double quantization |
| LoRA target modules | `all-linear` |
| LoRA rank | 64 |
| LoRA alpha | 16 |
| LoRA dropout | 0.05 |
| Max length | 2048 |
| Loss | assistant-only SFT loss |
| Training split | `train_sota` |
| Learning rate | 2e-4 |
| Scheduler | constant |
| Optimizer | paged AdamW 32-bit |
| Hardware | NVIDIA RTX 6000 Ada 48/50GB |
| Framework | TRL `SFTTrainer` + PEFT |

The dataset was audited with the Qwen3 chat template; all source rows fit within 2048 tokens.

## Main evaluation results

Stage-1 raw metrics:

| Split | JSON parse | Exact match | Field F1 | KPI presence |
|---|---:|---:|---:|---:|
| `test_in_distribution` | 1.0000 | 0.0227 | 0.6868 | 0.7973 |
| `test_template_ood` | 1.0000 | 0.0014 | 0.6790 | 0.8062 |
| `test_use_case_ood` | 0.9998 | 0.0122 | 0.6825 | 0.7883 |
| `test_sector_ood` | 1.0000 | 0.0166 | 0.6610 | 0.7733 |
| `test_adversarial` | 1.0000 | 0.9697 | 0.9697 | 1.0000 |

Stage-1 normalized metrics:

| Split | JSON parse | Normalized field F1 | Normalized key F1 | Normalized exact |
|---|---:|---:|---:|---:|
| `test_in_distribution` | 1.0000 | 0.7956 | 0.9811 | 0.0351 |
| `test_template_ood` | 1.0000 | 0.7865 | 0.9801 | 0.0177 |
| `test_use_case_ood` | 0.9998 | 0.7907 | 0.9805 | 0.0253 |
| `test_sector_ood` | 1.0000 | 0.7697 | 0.9818 | 0.0293 |
| `test_adversarial` | 1.0000 | 0.9697 | 1.0000 | 0.9697 |

Normalization removes volatile/generated fields such as IDs, hrefs, timestamps, schema links, descriptions, and generated identifiers before computing field/key metrics.

## Layer-level findings

Strong layers:

- `tmf921`: normalized field F1 around 0.93–0.94.
- `camara`: normalized field F1 around 0.81–0.87.
- `intent_3gpp`: normalized field F1 around 0.80–0.82.
- `etsi_zsm`: normalized field F1 around 0.75–0.79.

Weak layers:

- `o1_nrm`: normalized field F1 around 0.39–0.40.
- `a1_policy`: normalized field F1 around 0.67–0.68.
- `tmf921_lifecycle_report`: normalized field F1 around 0.15–0.18.
- `tmf921_lifecycle_monitor`: normalized field F1 around 0.39–0.52.

A stage-2 weak-layer continuation experiment was performed but is **not promoted** because it did not materially improve O1/A1 and slightly reduced adversarial robustness. Stage 1 remains the primary model.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "Qwen/Qwen3-8B"
adapter = "nraptisss/Qwen3-8B-TMF921-Intent-QLoRA-qwen3-8b-qlora-20260501-083834"

tokenizer = AutoTokenizer.from_pretrained(adapter, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

messages = [
    {"role": "system", "content": "You are a telecom intent-to-configuration assistant. Return valid JSON only."},
    {"role": "user", "content": "Create a URLLC slice for remote robotic surgery in Hospital Campus with 1 ms latency, 99.9999% reliability, 100 Mbps downlink, 50 Mbps uplink, and 20 UEs."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=1536, do_sample=False)
print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```

## Limitations

- This model is trained on synthetic research data, not real operator logs.
- It is not certified against official TMF921/3GPP/ETSI/CAMARA/O-RAN schemas.
- It should not be used to deploy real network configurations without expert review and validators.
- O1 NRM and A1 policy value fidelity remain open challenges.
- Raw exact match is low because many outputs contain volatile/generated fields.
- Normalized metrics are a research proxy, not proof of production standards compliance.

## Recommended citation

```bibtex
@model{raptis_qwen3_tmf921_qlora_2026,
  title = {Qwen3-8B TMF921 Intent-to-Configuration QLoRA Adapter},
  author = {Raptis, Nikolaos},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/nraptisss/Qwen3-8B-TMF921-Intent-QLoRA-qwen3-8b-qlora-20260501-083834}
}
```

## Related resources

- Dataset: https://huggingface.co/datasets/nraptisss/TMF921-intent-to-config-research-sota
- Training/evaluation repo: https://huggingface.co/nraptisss/tmf921-intent-training
- Results: https://huggingface.co/nraptisss/tmf921-intent-training/tree/main/results
- Qualitative examples: https://huggingface.co/nraptisss/tmf921-intent-training/tree/main/analysis