Update README.md

README.md

@@ -1,3 +1,169 @@
 ---
-license: apache-2.0
+language:
+- en
+library_name: peft
+pipeline_tag: text-generation
+base_model: zai-org/GLM-4.7-Flash
+base_model_relation: adapter
+tags:
+- cybersecurity
+- phishing-detection
+- email-security
+- glm
+- lora
+- peft
 ---
+
+# Gl00be Phishing Guard
+
+A specialized language-model adapter for defensive email phishing classification, developed by Gl00be.
+
+Built on GLM-4.7-Flash, Phishing Guard uses parameter-efficient post-training to adapt the model to security-focused email analysis.
+
+**Release status:** adapter files and inference implementation will be uploaded when ready.
+
+## Architecture
+
+| Component | Specification |
+|---|---|
+| Base model | zai-org/GLM-4.7-Flash |
+| Model family | GLM |
+| Architecture | Mixture-of-experts causal language model |
+| Total parameters | Approximately 30 billion |
+| Active parameters per token | Approximately 3 billion |
+| Adaptation method | Attention-only LoRA |
+| Trainable adapter parameters | 10,515,968 |
+| Trainable proportion | Approximately 0.0351% |
+| Base-weight precision during training | BF16 |
+| Implementation | PyTorch, Transformers and PEFT |
+
+The original model weights remain frozen during post-training. The adapter introduces trainable low-rank updates to selected attention projections.
+
+## Post-training
+
+Phishing Guard uses supervised fine-tuning on labeled English email examples.
+
+The training objective is completion-only causal language modeling: prompt tokens are masked from the loss, and the model learns to produce the classification label and end-of-sequence token.
+
+### LoRA configuration
+
+| Parameter | Value |
+|---|---|
+| Rank | 8 |
+| Alpha | 16 |
+| Dropout | 0.05 |
+| Bias adaptation | None |
+| Task type | CAUSAL_LM |
+
+Targeted attention modules:
+
+- `q_a_proj`
+- `q_b_proj`
+- `kv_a_proj_with_mqa`
+- `kv_b_proj`
+- `o_proj`
+
+### Training configuration
+
+| Parameter | Configuration |
+|---|---|
+| Training hardware | NVIDIA DGX Spark / GB10 |
+| Maximum sequence length | 1,024 tokens |
+| Microbatch size | 1 |
+| Gradient accumulation | 16 |
+| Effective batch size | 16 |
+| Initial learning rate | 0.0001 |
+| Optimizer | AdamW |
+| Planned training duration | 1 epoch |
+| Gradient checkpointing | Enabled |
+| Random seed | 42 |
+
+Training data preparation removes normalized exact duplicates and entries with conflicting labels before creating stratified training, validation and test splits.
+
+Dataset source: `zefang-liu/phishing-email-dataset`.
+
+Source revisions and preprocessing details will accompany the release.
+
+## Input and output
+
+### Input
+
+Email text supplied as untrusted content within a chat prompt.
+
+The system instruction directs the model to classify the email rather than follow instructions embedded within it. This prompt design does not establish immunity to prompt injection.
+
+Inputs are truncated to fit the configured token budget. Long messages may require application-level segmentation.
+
+### Output
+
+The expected response is one label:
+
+| Label | Meaning |
+|---|---|
+| `PHISHING` | The supplied text is classified as phishing |
+| `LEGITIMATE` | No phishing detected in the supplied text |
+
+Applications should handle any other output as an invalid classification requiring review.
+
+`LEGITIMATE` is not a guarantee that an email, sender, URL or attachment is safe.
+
+## Reuse and integration
+
+The planned release distributes a PEFT LoRA adapter rather than a standalone copy of the full base model.
+
+Local inference requires:
+
+1. Loading the compatible GLM-4.7-Flash base-model revision.
+2. Loading the tokenizer and adapter.
+3. Applying the documented chat template and preprocessing.
+4. Generating and validating the classification label.
+
+The release will include the adapter configuration, adapter weights, tokenizer files, base-model revision and tested inference instructions.
+
+### Planned repository contents
+
+| File | Purpose |
+|---|---|
+| `adapter_config.json` | PEFT configuration and base-model reference |
+| `adapter_model.safetensors` | Trained LoRA weights |
+| Tokenizer files | Tokenization and chat-template configuration |
+| `run.json` | Training configuration and base-model revision |
+| `inference.py` | Local inference entry point |
+| Dependency specification | Compatible runtime packages |
+| `README.md` | Model documentation and usage instructions |
+
+These files will be added when the release is ready.
+
+## Local deployment
+
+Phishing Guard is intended for integration into email triage, security tools and defensive agent workflows.
+
+Users can load the base model and adapter within their own Python applications using Transformers and PEFT.
+
+The adapter is small relative to the base model, but inference still requires sufficient memory for the underlying GLM model. Quantized deployment requires a compatible implementation and separate verification.
+
+## Operational scope
+
+This version performs text-based binary classification.
+
+It does not independently:
+
+- Verify SPF, DKIM or DMARC.
+- Establish sender identity.
+- Inspect binary attachments.
+- Retrieve live URL reputation.
+- Provide calibrated confidence scores.
+
+Application integrations should combine classifications with email metadata, established security controls and human review where appropriate.
+
+## Licensing and attribution
+
+GLM-4.7-Flash is the upstream base model developed by Z.ai.
+
+The base model and training dataset have separate licensing terms. Adapter release terms and attribution notices will be provided with the downloadable files.
+
+## About Gl00be
+
+Gl00be develops AI cybersecurity systems focused on autonomous agents, vulnerability discovery and defensive intelligence.
+
+Website: https://cybersecurity.gl00be.com