docs: document fine-tuning datasets and training hardware in model card

README.md

@@ -138,6 +138,73 @@ The plugin fuses these model probabilities with its regex presets and custom
 patterns; regex `increment` deltas raise a question's probability and can
 escalate the action. See the plugin's `.planning/ARCHITECTURE.md`.
 
+## Training datasets
+
+This checkpoint (v3) was fine-tuned on a mixture of public prompt-safety and PII
+corpora plus curated counter-examples: **32,256 unique states** (29,037 train /
+3,219 validation), expanded to **82,582 per-question training items** and 800
+calibration items. Sources were selected for reliability, clear licensing, and
+coverage of the three target risk families.
+
+| Question(s) | Dataset | Rows | License |
+| --- | --- | --- | --- |
+| `injection_present` | [`deepset/prompt-injections`](https://huggingface.co/datasets/deepset/prompt-injections) | 546 | CC-BY-4.0 |
+| `injection_present` | [`Lakera/mosscap_prompt_injection`](https://huggingface.co/datasets/Lakera/mosscap_prompt_injection) | 4,319 | MIT |
+| `injection_present` | [`xTRam1/safe-guard-prompt-injection`](https://huggingface.co/datasets/xTRam1/safe-guard-prompt-injection) | 4,598 | see card |
+| `injection_present` | [`neuralchemy/Prompt-injection-dataset`](https://huggingface.co/datasets/neuralchemy/Prompt-injection-dataset) | 4,261 | Apache-2.0 |
+| `jailbreak_attempt`, `injection_present` | [`TrustAIRLab/in-the-wild-jailbreak-prompts`](https://huggingface.co/datasets/TrustAIRLab/in-the-wild-jailbreak-prompts) (`jailbreak` + `regular` configs) | 653 + 5,484 | MIT |
+| `harmful_request` | [`JailbreakBench/JBB-Behaviors`](https://huggingface.co/datasets/JailbreakBench/JBB-Behaviors) (`harmful` + `benign` splits) | 100 + 100 | MIT |
+| `harmful_request` | [`AlignmentResearch/HarmBench`](https://huggingface.co/datasets/AlignmentResearch/HarmBench) | 200 | see card |
+| `harmful_request`, `jailbreak_attempt` | [`lmsys/toxic-chat`](https://huggingface.co/datasets/lmsys/toxic-chat) | 2,993 | CC-BY-NC-4.0 |
+| `pii_*` | [`ai4privacy/pii-masking-400k`](https://huggingface.co/datasets/ai4privacy/pii-masking-400k) | 2,998 | other (see card) |
+| benign negatives | [`databricks/databricks-dolly-15k`](https://huggingface.co/datasets/databricks/databricks-dolly-15k) | 2,978 | CC-BY-SA-3.0 |
+| benign negatives | [`HuggingFaceH4/no_robots`](https://huggingface.co/datasets/HuggingFaceH4/no_robots) | 2,997 | CC-BY-NC-4.0 |
+| counter-examples | curated `hard_negatives` (this project) | 29 | — |
+
+Per-question label counts after expansion: `injection_present` 9,016 positive /
+16,849 negative; `pii_present` 1,500 / 7,502; `pii_person_name` 1,071 / 1,927;
+`pii_address` 1,168 / 1,830; `pii_ssn` 650 / 2,348; `pii_email` 431 / 2,567;
+`pii_phone` 505 / 2,493; `pii_secret` 135 / 2,863; `pii_credit_card` 86 /
+2,912; `jailbreak_attempt` 734 / 14,400; `harmful_request` 594 / 8,803.
+`pii_ip_address` has no positive training signal in this corpus (IP addresses
+are rare in `ai4privacy`); the consumer's regex preset handles it
+deterministically.
+
+Rows are deduplicated on `sha1(state)` across all sources and split per source
+90/10 (seed `20261004`). Each training row carries **partial gold** — it only
+labels the questions its source can answer. No training data is redistributed
+in this repository; only the fine-tuned weights are published. Refer to each
+dataset card for its own terms.
+
+## Training hardware & configuration
+
+Trained with the `opencode-bunker-laya` pipeline on a single consumer GPU.
+
+| | |
+| --- | --- |
+| GPU | NVIDIA GeForce RTX 5060 Ti, 16 GB (Blackwell, compute capability 12.0 / `sm_120`) |
+| GPU (smoke tests) | NVIDIA GeForce RTX 4060 Ti, 8 GB (`sm_89`) |
+| NVIDIA driver | 580.173.02 |
+| CUDA / PyTorch | CUDA 13.0, PyTorch 2.14.1+cu130 |
+| Precision | bf16 autocast |
+| Parallelism | single device (`world_size = 1`, one GPU) |
+
+Key hyperparameters:
+
+| Setting | Value |
+| --- | --- |
+| Base model | `convaiinnovations/laya` (ModernBERT-large, 421M) |
+| Objective | RLCD (strictly proper scoring rule, `w_sph = 0.75`, `w_rps = 1.0`) + cross-entropy, GRPO-style advantage normalisation |
+| Epochs | 2 (v3 run) |
+| Micro-batch / grad-accum | 4 / 16 (effective batch 64) |
+| Optimiser | AdamW, encoder LR `2.5e-5`, head LR `1.0e-4`, cosine schedule, grad clip 1.0 |
+| Max sequence / head length | 1,024 / 256 |
+| Gradient checkpointing | on |
+| Calibration | one temperature per question type (`[1.2, 1.2, 4.433]`) |
+
+Final results: macro accuracy **0.987** across the 12 questions. Per-question
+accuracy, Brier score and ECE are in `eval_metrics.json`.
+
 ## Provenance
 
 Built by the `opencode-bunker-laya` pipeline. Base model Apache-2.0. Each