Updated model card

README.md

@@ -1,13 +1,13 @@
 # Ginie — Smart Contract LLM
 
-The first AI model purpose-built to generate, compile, audit, and deploy smart contracts across institutional and public blockchains. Plain English in. Production-ready contract out. On-chain in under 90 seconds.
-
-[![Website](https://img.shields.io/badge/Website-ginie.xyz-blue)](https://ginie.xyz)
 [![npm](https://img.shields.io/badge/npm-30k%2B_weekly_downloads-red)](https://npmjs.com/package/ginie-sdk)
+[![Website](https://img.shields.io/badge/Website-ginie.xyz-blue)](https://ginie.xyz)
 [![License](https://img.shields.io/badge/License-MIT-green)](https://opensource.org/licenses/MIT)
 [![Demo](https://img.shields.io/badge/Demo-Live-brightgreen)](https://huggingface.co/spaces/GinieAI/Ginie-Demo)
 [![Canton](https://img.shields.io/badge/Canton_Network-Supported-purple)](https://canton.network)
 
+The first AI model purpose-built to generate, compile, audit, and deploy smart contracts across institutional and public blockchains. Plain English in. Production-ready contract out. On-chain in under 90 seconds.
+
 ---
 
 ## What is Ginie?
@@ -16,12 +16,11 @@ Ginie is the developer layer for the next generation of on-chain applications. T
 
 Write a description. Get a contract that compiles, passes security checks, and deploys — across Solidity (Ethereum, Avalanche, Camp Network), Daml (Canton Network), and Rust (Vara Network).
 
-Canton Network alone processes $6 trillion in tokenised assets, backed by Goldman Sachs, JPMorgan, and DTCC. Every institution building on it needs smart contracts. Ginie writes them.
+Canton Network processes $6 trillion in tokenised assets, backed by Goldman Sachs, JPMorgan, and DTCC. Every institution building on it needs smart contracts. Ginie writes them.
 
 ---
 
 ## Quickstart
-
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
@@ -30,7 +29,7 @@ tokenizer = AutoTokenizer.from_pretrained("GinieAI/Solidity-LLM")
 model = AutoModelForCausalLM.from_pretrained(
     "GinieAI/Solidity-LLM",
     torch_dtype=torch.bfloat16,
-    device_map="cuda"
+    device_map="auto"
 )
 
 prompt = """### Instruction:
@@ -39,7 +38,7 @@ Write a Solidity ERC20 token contract with minting, burning, and owner controls.
 ### Response:
 """
 
-inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
 outputs = model.generate(
     **inputs,
     max_new_tokens=800,
@@ -49,13 +48,13 @@ outputs = model.generate(
 )
 
 print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
 
-
-npm SDK — 30,000+ active weekly downloads
-
+**npm SDK** — 30,000+ active weekly downloads
+```bash
 npm install ginie-sdk
-
-
+```
+```javascript
 import { Ginie } from 'ginie-sdk'
 
 const ginie = new Ginie({ apiKey: 'your-key' })
@@ -69,85 +68,176 @@ const contract = await ginie.generate({
 console.log(contract.code)
 console.log(contract.securityScore)
 console.log(contract.compiled)
+```
 
+---
 
-Model Details
+## Model Details
 
+### Model Description
 
+- **Developed by:** [Ginie AI](https://ginie.xyz)
+- **Model type:** Causal LM — Code Generation
+- **Language:** English instructions → Solidity / Daml
+- **Parameters:** 2 Billion
+- **Architecture:** 32 Transformer blocks
+- **Context length:** 2048 tokens
+- **Precision:** bfloat16
+- **Tokenizer:** GPT2Tokenizer
+- **Finetuned from:** [Chain-GPT/Solidity-LLM](https://huggingface.co/Chain-GPT/Solidity-LLM)
+- **License:** MIT
 
-|Property      |Value                        |
-|--------------|-----------------------------|
-|Developer     |[Ginie AI](https://ginie.xyz)|
-|Model type    |Causal LM — Code Generation  |
-|Parameters    |2 Billion                    |
-|Architecture  |32 Transformer blocks        |
-|Context length|2048 tokens                  |
-|Precision     |bfloat16                     |
-|Tokenizer     |GPT-2                        |
-|Base model    |Chain-GPT/Solidity-LLM       |
-|License       |MIT                          |
+### Model Sources
 
-Performance
-Evaluated against GPT-4o mini and DeepSeek-Coder-7B on 100 Solidity contract generation prompts. Compilation success and security assessed via Slither static analysis. OpenZeppelin compliance assessed against standard library usage patterns.
+- **Demo:** [huggingface.co/spaces/GinieAI/Ginie-Demo](https://huggingface.co/spaces/GinieAI/Ginie-Demo)
+- **Website:** [ginie.xyz](https://ginie.xyz)
+- **npm SDK:** [npmjs.com/package/ginie-sdk](https://npmjs.com/package/ginie-sdk)
 
+---
 
+## Uses
 
-|Metric                 |Ginie v1|GPT-4o mini|DeepSeek-Coder-7B|
-|-----------------------|--------|-----------|-----------------|
-|Compilation rate       |**83%** |78%        |75%              |
-|OpenZeppelin compliance|**65%** |61%        |58%              |
-|Gas efficiency         |**72%** |65%        |63%              |
-|Security score         |**58%** |54%        |51%              |
+### Direct Use
 
-Ginie achieves the highest compilation rate despite being the smallest model in the comparison — a direct result of domain specialisation over general-purpose scale.
+- ERC20, ERC721, ERC1155 token contracts
+- DeFi protocols — staking, liquidity pools, yield farming
+- DAO and governance contracts
+- Multisig wallets and escrow agreements
+- NFT marketplaces
+- Automated compliance and audit loops
 
-What Ginie generates today
-	∙	ERC20, ERC721, ERC1155 token contracts
-	∙	DeFi protocols — staking, liquidity pools, yield farming
-	∙	DAO and governance contracts
-	∙	Multisig wallets and escrow agreements
-	∙	NFT marketplaces
-	∙	Automated compliance and audit loops
-Chains supported
+### Downstream Use
 
+- Integrated into IDEs and smart contract development platforms
+- Embedded in agentic pipelines for autonomous contract deployment
+- npm SDK for direct integration in any JavaScript or TypeScript project
 
+### Out-of-Scope Use
 
-|Chain         |Language|Status    |
-|--------------|--------|----------|
-|Ethereum      |Solidity|Live      |
-|Avalanche     |Solidity|Live      |
-|Camp Network  |Solidity|Live      |
-|Canton Network|Daml    |v3 roadmap|
-|Vara Network  |Rust    |v3 roadmap|
+- Production deployment without expert review
+- Formal legal or compliance auditing
+- Non-code generation tasks
 
-Not suitable for
-	∙	Production deployment without expert review
-	∙	Formal legal contract auditing
-	∙	Non-code generation tasks
+### Bias, Risks, and Limitations
 
-Roadmap
+- May reflect patterns from web-scraped Solidity including outdated or insecure practices
+- Can generate syntactically valid but logically incorrect contracts
+- Security score is indicative — not a substitute for a professional audit
+- All generated contracts must be reviewed before mainnet deployment
 
+---
 
+## How to Get Started
 
-|Version   |What ships                                               |
-|----------|---------------------------------------------------------|
-|v1.0 (now)|Solidity generation — 2B params, 83% compile rate        |
-|v2.0      |Expanded corpus — DISL + Zellic, 7,800+ training examples|
-|v3.0      |Daml and Rust support — Canton Network and Vara Network  |
-|v4.0      |Data flywheel — weekly retraining on real user prompts   |
+See the [Quickstart](#quickstart) section above. For streaming output:
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
+import torch
 
-The v4 flywheel is the permanent moat. Every contract a user successfully generates becomes a training example for the next version. The model improves weekly from real usage — a data distribution no statically trained competitor can replicate.
+tokenizer = AutoTokenizer.from_pretrained("GinieAI/Solidity-LLM")
+model = AutoModelForCausalLM.from_pretrained(
+    "GinieAI/Solidity-LLM",
+    torch_dtype=torch.bfloat16,
+    device_map="auto"
+)
 
-Training
-Ginie v1 is fine-tuned from Chain-GPT/Solidity-LLM using LoRA on a curated Solidity instruction dataset. Training focused on instruction-following quality, OpenZeppelin pattern adherence, and compilable output over raw token prediction.
-Security validation uses Slither static analysis. Compilation validation uses solc. Both are integrated into the generation pipeline — not just evaluation.
+prompt = """### Instruction:
+Write a Solidity staking contract with reward distribution.
 
-License and Attribution
-Released under the MIT License.
-Built on Chain-GPT/Solidity-LLM by ChainGPT, which is fine-tuned from Salesforce/codegen-2B-multi. Full credit to the original authors. Ginie extends this work for the institutional blockchain ecosystem.
+### Response:
+"""
+
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+
+model.generate(
+    **inputs,
+    max_new_tokens=1000,
+    temperature=0.7,
+    do_sample=True,
+    streamer=streamer,
+    pad_token_id=tokenizer.eos_token_id
+)
+```
+
+---
+
+## Training Details
+
+### Training Data
+
+- [AlfredPros/smart-contracts-instructions](https://huggingface.co/datasets/AlfredPros/smart-contracts-instructions) — 6,000 instruction-output pairs
+- Filtered: Solidity >= 0.5, token length 200–4000, compilable contracts only
+- Deduplicated and validated before training
+
+### Training Procedure
+
+- **Method:** LoRA fine-tuning (rank 16, alpha 16)
+- **Base:** Chain-GPT/Solidity-LLM → Salesforce/codegen-2B-multi
+- **Epochs:** 3
+- **Precision:** bfloat16
+- **Framework:** Unsloth + HuggingFace TRL
+- **Validation:** Slither static analysis + solc compilation check
+
+---
+
+## Evaluation
+
+Evaluated on 100 held-out Solidity prompts. Compilation measured via `solc`. Security measured via [Slither](https://github.com/crytic/slither).
+
+| Metric | Ginie v1 | GPT-4o mini | DeepSeek-Coder-7B |
+|---|---|---|---|
+| Compilation rate | **83%** | 78% | 75% |
+| OpenZeppelin compliance | **65%** | 61% | 58% |
+| Gas efficiency | **72%** | 65% | 63% |
+| Security score | **58%** | 54% | 51% |
+
+Ginie achieves the highest compilation rate at 2B parameters — domain specialisation over general-purpose scale.
+
+---
+
+## Chains Supported
+
+| Blockchain | Language | Status |
+|---|---|---|
+| Ethereum | Solidity | Live |
+| Avalanche | Solidity | Live |
+| Camp Network | Solidity | Live |
+| Canton Network | Daml | v3 roadmap |
+| Vara Network | Rust | v3 roadmap |
+
+---
+
+## Roadmap
+
+| Version | What ships |
+|---|---|
+| v1.0 (current) | Solidity generation — 2B params, 83% compile rate |
+| v2.0 | Expanded corpus — DISL + Zellic, 7,800+ examples, security score 70%+ |
+| v3.0 | Daml + Rust — Canton Network and Vara Network |
+| v4.0 | Weekly retraining flywheel on verified user prompts |
+
+---
+
+## Citation
+```bibtex
+@misc{ginie2025,
+  title     = {Ginie: Smart Contract LLM for Institutional Blockchain},
+  author    = {Ginie AI},
+  year      = {2025},
+  url       = {https://huggingface.co/GinieAI/Solidity-LLM},
+  note      = {Fine-tuned from Chain-GPT/Solidity-LLM (Salesforce/codegen-2B-multi)}
+}
+```
+
+---
+
+## License and Attribution
+
+Released under the MIT License. Built on [Chain-GPT/Solidity-LLM](https://huggingface.co/Chain-GPT/Solidity-LLM) by ChainGPT, fine-tuned from [Salesforce/codegen-2B-multi](https://huggingface.co/Salesforce/codegen-2B-multi). Full credit to the original authors.
+
+---
 
-About Ginie AI
-Ginie AI is building the developer layer for institutional blockchain. Backed by the Canton Foundation and supported by the Canton Network ecosystem — the institutional blockchain processing $6 trillion in tokenised assets with Goldman Sachs, JPMorgan, and DTCC.
-ginie.xyz · npm SDK · Live demo
+[ginie.xyz](https://ginie.xyz) · [npm SDK](https://npmjs.com/package/ginie-sdk) · [Live demo](https://huggingface.co/spaces/GinieAI/Ginie-Demo)
 
-Smart contracts generated by Ginie should be reviewed by a qualified developer before production deployment. Security scores are indicative and do not constitute a formal audit.
+> Smart contracts generated by Ginie require expert review before production deployment. Security scores are indicative and do not constitute a formal audit.