Add research guide: 9 research directions with concrete experiments

RESEARCH_GUIDE.md

@@ -0,0 +1,157 @@
+# Research Guide: How TelecomIntentBench Enables Intent Translation Research
+
+## 1. Fine-Tuning Open-Weight Intent Translators
+
+The most direct use. Today, ORION (arxiv:2603.03667) achieves 100% policy generation success — but only with proprietary frontier models (GPT-5, Claude Opus 4.5). There is no open-weight model that can do this. The dataset lets you change that:
+
+```python
+from trl import SFTTrainer, SFTConfig
+from datasets import load_dataset
+
+ds = load_dataset("nraptisss/telecom-intent-config-sft-10k")
+
+trainer = SFTTrainer(
+    model="Qwen/Qwen2.5-7B-Instruct",
+    train_dataset=ds["train"],
+    eval_dataset=ds["test"],
+    args=SFTConfig(output_dir="./telecom-intent-7b", num_train_epochs=3)
+)
+```
+
+**Research question:** *Can a 7B open model match GPT-5's intent translation accuracy when fine-tuned on schema-grounded data?* If yes, this removes the dependency on proprietary APIs for autonomous network orchestration — a critical requirement for edge/RAN deployment where cloud API calls add unacceptable latency.
+
+---
+
+## 2. Multi-Layer Translation Chain Experiments
+
+No existing dataset covers more than one configuration layer. TelecomIntentBench covers four — the same four layers that a real intent traverses in production:
+
+```
+Operator NL → CAMARA booking → 3GPP intent → A1 policy → O1 NRM config
+```
+
+This enables **cascaded translation** research:
+
+- **Single-model generalist:** Train one model on all four layers, steer via system prompt. Measure whether a single model degrades vs. specialists.
+- **Four specialist pipeline:** Train four layer-specific models (filter by `target_layer`), chain them. The output of model₁ (CAMARA JSON) becomes the input context for model₂ (A1 policy), etc.
+- **Cross-layer consistency verification:** Given the same underlying KPI targets, do the four layer outputs agree? The metadata columns (`latency_ms`, `dl_throughput_mbps`, etc.) provide ground truth for cross-layer consistency checks.
+
+**Research question:** *Does a cascaded specialist pipeline produce more schema-valid and internally consistent end-to-end translations than a single generalist model?*
+
+---
+
+## 3. Schema-Constrained Decoding Research
+
+A core problem in intent translation is **hallucinated configurations** — the model invents network parameters that don't exist (e.g., a `prbQuotaMin` of 150%, or a nonexistent `cellId`). The Microsoft NeMo paper (arxiv:2308.06261) found that 9/35 failures were due to "imaginary graph attributes."
+
+TelecomIntentBench provides the schemas needed to build and evaluate **constrained decoding** approaches:
+
+- Use the CAMARA/A1/3GPP/O1 schemas as JSON Schema constraints during generation (e.g., via `outlines`, `guidance`, or grammar-constrained sampling)
+- The 100% schema compliance of the training data teaches the model the valid output space
+- The test split (1,000 samples) provides a benchmark: measure **schema validity rate** and **field accuracy** on held-out intents
+- Compare: unconstrained generation vs. schema-constrained decoding vs. MCP tool-use (ORION's approach)
+
+**Research question:** *Which approach — fine-tuning, constrained decoding, or tool-use — yields the highest schema compliance at the lowest inference latency?*
+
+---
+
+## 4. Robustness and Ambiguity Studies
+
+The dataset's current intents are well-specified (every KPI is explicit). Real operator intents are not. This creates a research opportunity:
+
+- **Ablation by omission:** Systematically remove KPI fields from test intents ("Provision a URLLC slice for remote surgery in Hospital Campus" — no latency, no throughput mentioned). Measure whether the fine-tuned model infers reasonable defaults from the slice type and use case context.
+- **Paraphrase robustness:** Use an LLM to rephrase the 1,000 test intents into colloquial language ("make the network fast enough for the surgeon to operate remotely"). Measure accuracy degradation.
+- **Adversarial intents:** Construct contradictory intents ("low-latency mMTC with 1M devices at 0.5ms" — physically impossible). Test whether the model flags conflicts or silently generates invalid configs.
+
+The metadata columns make these experiments trivial to set up — you know the ground truth KPIs for every sample.
+
+**Research question:** *How gracefully does a fine-tuned intent translator degrade when intents are ambiguous, underspecified, or contradictory?*
+
+---
+
+## 5. Telecom LLM Benchmarking
+
+The test split serves as a standardized **intent translation benchmark** with clear metrics:
+
+| Metric | What it measures |
+|---|---|
+| **JSON validity rate** | % of outputs that parse as valid JSON |
+| **Schema compliance rate** | % that contain all required fields for the target layer |
+| **Field accuracy** | Exact match on categorical fields (sliceType, sst, sd) |
+| **KPI deviation** | Mean absolute error on numeric fields vs. ground truth |
+| **Cross-layer consistency** | Agreement between layer outputs for the same intent |
+
+You can benchmark any model — ORANSight variants, TelecomGPT, general-purpose LLMs, or your own fine-tuned models — on the same 1,000 test intents. Currently no such standardized intent translation benchmark exists (ORAN-Bench-13K and TeleQnA are MCQ, not generative).
+
+---
+
+## 6. Data Augmentation and Curriculum Learning
+
+The dataset's controlled structure enables principled data augmentation research:
+
+- **Curriculum by complexity:** Start training on CAMARA (simplest schema, ~430 chars output), progress to 3GPP intent (~1,200 chars), then A1 policy, then O1 NRM. Measure whether curriculum ordering improves convergence.
+- **Slice-type curriculum:** Start with eMBB (relaxed KPIs), progress to URLLC (tight constraints). Hypothesis: models that learn relaxed constraints first generalize better to strict ones.
+- **Synthetic augmentation scaling laws:** Generate 10K, 50K, 100K, 500K samples using the released pipeline. Plot accuracy vs. dataset size. Find the saturation point — how much synthetic data is enough before you need real operator data?
+
+---
+
+## 7. Continual Pre-Training on Structured Telecom Data
+
+Following the TelecomGPT pipeline (arxiv:2407.09424), the ~3.85M tokens of structured JSON in TelecomIntentBench can be used as **domain-specific continual pre-training data** — not for the intent task directly, but to teach the model the vocabulary and structure of telecom configurations (S-NSSAI, PRB ratios, A1 policies, NRCellDU hierarchies). This is complementary to the OpenTelecom corpus (which is unstructured 3GPP spec text).
+
+**Research question:** *Does continual pre-training on structured config JSON improve downstream intent translation accuracy more than pre-training on the same volume of unstructured 3GPP spec text?*
+
+---
+
+## 8. DPO/RLHF Alignment for Telecom
+
+The dataset provides the SFT stage. The natural next step is **alignment**:
+
+- Generate multiple candidate configs for each test intent using the SFT model
+- Validate each candidate against the schema → schema-valid = "chosen", schema-invalid = "rejected"
+- Train DPO on these preference pairs
+
+TelecomGPT showed that DPO gives slight gains on math modeling but slight drops on MCQ. The question for intent translation is different — DPO should directly optimize for schema compliance, which is the metric that matters.
+
+**Research question:** *Does DPO alignment on schema-valid/invalid preference pairs improve schema compliance rates beyond what SFT achieves alone?*
+
+---
+
+## 9. Edge Deployment and Efficiency Research
+
+6G-Bench (arxiv:2602.08675) showed that model scale ≠ monotone improvement on telecom reasoning tasks, and that MoE models are competitive at lower active parameter counts. TelecomIntentBench enables:
+
+- **QLoRA efficiency studies:** Following TSLAM-Mini's recipe (LoRA r=16, α=32, NF4), fine-tune models from 1.5B to 32B and plot accuracy vs. model size. Find the minimum viable model for each layer.
+- **Quantization impact:** Compare FP16, INT8, INT4 inference on intent translation accuracy. The structured JSON output is particularly sensitive to quantization errors (a single misplaced digit changes the entire config).
+- **Latency-accuracy tradeoff:** For real-time intent orchestration (ORION targets seconds), measure end-to-end latency for local model inference vs. cloud API calls.
+
+---
+
+## Research Roadmap
+
+| Phase | What to do | What you learn |
+|---|---|---|
+| **Phase 1** | SFT on the dataset, evaluate on test split | Baseline accuracy for open models |
+| **Phase 2** | Compare generalist vs. specialist vs. cascade | Best architecture for multi-layer translation |
+| **Phase 3** | Robustness testing with ambiguous/adversarial intents | Failure modes and safety boundaries |
+| **Phase 4** | DPO alignment on schema-valid preferences | Whether alignment improves schema compliance |
+| **Phase 5** | Scaling laws (10K→500K) + model size sweeps | Minimum data and model size for production use |
+| **Phase 6** | Integration with ORION-like MCP architecture | End-to-end autonomous orchestration without proprietary models |
+
+---
+
+## Key References
+
+| Paper | Relevance |
+|---|---|
+| [ORION](https://arxiv.org/abs/2603.03667) (2025) | CAMARA schema, A1 policy format, 100% success with frontier models |
+| [TelecomGPT](https://arxiv.org/abs/2407.09424) (2024) | 3-stage pipeline (CPT→SFT→DPO), continual pre-training methodology |
+| [ORANSight 2.0](https://arxiv.org/abs/2503.05200) (2025) | RANSTRUCT synthetic data generation, QLoRA fine-tuning |
+| [TSLAM-Mini](https://arxiv.org/abs/2505.07877) (2025) | QLoRA training config, edge deployment viability |
+| [NEFMind](https://arxiv.org/abs/2508.09240) (2025) | Telecom API automation, synthetic data from API specs |
+| [6G-Bench](https://arxiv.org/abs/2602.08675) (2026) | Scale vs. efficiency findings, MoE competitiveness |
+| [NeMo/NeMoEval](https://arxiv.org/abs/2308.06261) (2023) | Hallucination analysis in network config generation |
+| 3GPP TS 28.312 | Intent-Driven Management formal data model |
+| 3GPP TS 28.541 | NR NRM YANG model for RAN configuration |
+| O-RAN WG2 A1 spec | A1 policy JSON format |
+| CAMARA NetworkSliceBooking | High-level slice booking API schema |