sync README with paper-final numbers (8× A100, 13 sources, 78,358 samples)

README.md

@@ -15,45 +15,90 @@ size_categories:
 
 # Beam Training Data
 
-Supervised fine-tuning (SFT) dataset used to train the TorchSight Beam model — a cybersecurity document classifier based on Qwen 3.5 27B.
+Supervised fine-tuning (SFT) dataset used to train **TorchSight Beam** — a
+cybersecurity document classifier built by LoRA-fine-tuning Qwen 3.5 27B.
+
+> Dobrovolskyi, I. *Security Document Classification with a Fine-Tuned Local
+> Large Language Model: Benchmark Data and an Open-Source System.* Journal of
+> Information Security and Applications, 2026.
 
 ## Dataset
 
-- **74,441 training samples** (`sft/train_alpaca.jsonl`)
-- **3,917 validation samples** (`sft/val_alpaca.jsonl`)
+- **78,358 balanced samples** (95 / 5 split → 74,441 train + 3,917 validation)
 - Alpaca format: `instruction`, `input`, `output`
-- Balanced across 7 categories + subcategories
-
-## Sources (all verified safe for AI training)
-
-| Source | License | Content |
-|---|---|---|
-| AI4Privacy (300K PII) | Apache 2.0 | PII samples |
-| Enron (FERC release) | Public domain | Email/financial data |
-| NVD/NIST | Public domain (US Gov) | Vulnerability descriptions |
-| SecLists | MIT | Security payloads |
-| PayloadsAllTheThings | MIT | Attack payloads |
-| Prompt Injection datasets | Apache 2.0 | Injection attacks |
-| GHSA | CC-BY 4.0 | Security advisories |
-| Loghub | Research-free | System logs (safe class) |
-| Synthetic | Generated | Hard negatives, edge cases |
+- Output is a JSON array of findings, each with `category`, `subcategory`,
+  `severity`, `explanation`
+- Stratified across 7 categories × 51 subcategories
 
-## Structure
+## Composition
+
+The initial corpus contained 116,956 raw samples drawn from 13 publicly
+available sources. To remove the NVD-dominated skew, each subcategory was
+capped at 5,000 samples and underrepresented subcategories were augmented with
+GPT-4–generated synthetic data and hard-negative boundary cases.
 
-- `sft/` — Final SFT training files (Alpaca format)
-- `processed/` — Intermediate processed files from each source
-- `synthetic/` — Generated synthetic data (hard negatives, edge cases)
+| Source                | Raw    | Balanced | License        | Categories               |
+|---|---:|---:|---|---|
+| NVD CVE Database      | 50,000 |  8,475 | Public domain  | `malicious.exploit`           |
+| Synthetic augmentation| 33,100 | 39,754 | Generated (GPT-4) | All categories               |
+| Hard negatives        |  6,400 |  6,134 | Generated (GPT-4) | Boundary cases               |
+| AI4Privacy            |  5,000 |  4,851 | Apache 2.0     | `pii.*`                       |
+| Fenrir v2.0           |  5,000 |  4,573 | Apache 2.0     | `malicious.*`                 |
+| SEC EDGAR             |  3,000 |  3,000 | Public domain  | `financial.*`                 |
+| SecLists              |  3,229 |  1,708 | MIT            | `malicious.injection`         |
+| Phishing Dataset      |  3,000 |  2,796 | Apache 2.0     | `malicious.phishing`          |
+| NIST Training         |  3,000 |  2,761 | Public domain  | `safe.documentation`          |
+| Enron Email Corpus    |  2,000 |  1,902 | Public domain  | `pii.*`, `credentials.*`      |
+| MITRE ATT&CK v14      |  1,620 |    871 | Royalty-free   | `malicious.malware`           |
+| Loghub                |  1,280 |  1,280 | Research-free  | `safe.config`                 |
+| Other (3 sources)     |    327 |    253 | Permissive     | Multiple                       |
+| **Total**             | **116,956** | **78,358** | | |
 
-## Training
+All sources have been verified safe for AI training. Copyleft-licensed
+(GPL/LGPL) and ShareAlike-licensed (CC BY-SA) materials are excluded so the
+corpus is suitable for commercial training use.
+
+Of the final 78,358 samples, 39,754 (50.7%) are GPT-4–generated synthetic
+augmentation and 6,134 (7.8%) are hard-negative boundary cases. The remaining
+32,470 (41.5%) come from the 13 external sources listed above.
+
+## Structure
 
-```bash
-# LoRA training on Qwen 3.5 27B
-python train_lora.py  # r=128, alpha=256, 5 epochs, H100 80GB
-python export_gguf.py  # Export to GGUF for Ollama
 ```
+sft/
+├── train_alpaca.jsonl   # 74,441 samples
+└── val_alpaca.jsonl     #  3,917 samples
+processed/               # intermediate per-source files
+synthetic/               # GPT-4 generations and hard negatives
+```
+
+## LoRA training configuration
+
+| Parameter             | Value |
+|---|---|
+| Base model            | Qwen 3.5 27B (dense) |
+| LoRA rank (r)         | 128 |
+| LoRA alpha (α)        | 256 |
+| Target modules        | q/k/v/o_proj, gate/up/down_proj |
+| Dropout               | 0.05 |
+| Learning rate         | 2 × 10⁻⁵, cosine decay, 10% warmup |
+| Effective batch size  | 16 (4 × 4 gradient accumulation) |
+| Epochs                | 5 |
+| Precision             | bf16 |
+| Max sequence length   | 4,096 tokens |
+| Hardware              | 8× NVIDIA A100 80GB SXM4 |
+| Wall-clock time       | 10.5 hours |
 
-Compatible: trl 0.11.4 + transformers 4.45.2 + peft 0.13.2
+Library versions: `trl == 0.11.4`, `transformers == 4.45.2`, `peft == 0.13.2`.
 
 ## License
 
-Apache 2.0
+Apache 2.0.
+
+## Companion artifacts
+
+- Benchmark: [`torchsight/cybersecurity-classification-benchmark`](https://huggingface.co/datasets/torchsight/cybersecurity-classification-benchmark)
+- Models: [`torchsight/beam-q4_K_M`](https://huggingface.co/torchsight/beam-q4_K_M),
+  [`torchsight/beam-q8_0`](https://huggingface.co/torchsight/beam-q8_0),
+  [`torchsight/beam-f16`](https://huggingface.co/torchsight/beam-f16)
+- Source: [github.com/IvanDobrovolsky/torchsight](https://github.com/IvanDobrovolsky/torchsight)