Update README.md

Browse files

Files changed (1) hide show

README.md +93 -92

README.md CHANGED Viewed

@@ -25,135 +25,136 @@ license_link: LICENSE
 ---
-# 🚀 JumpLander Coder 32B
-**Advanced Code-Generation LLM — optimized for English & Persian workflows**
-🇮🇷 *چند خط توضیح فارسی:*
-این مدل برای توسعه‌دهندگان ایرانی طراحی شده و نسخه آنلاین آن در سایت فعال است.
-نسخه‌ی لوکال فقط از طریق نرم‌افزار رسمی JumpLander ارائه خواهد شد.
-وزن‌های مدل عمومی نیستند و تنها در قالب نرم‌افزار قابل استفاده می‌باشند.
----
-## 🌟 Overview
-JumpLander Coder 32B is a high-performance, bilingual (English–Persian) code-generation LLM built for advanced programming tasks, repository-wide reasoning, and architecture-level understanding.
-This release provides documentation, benchmarks, design goals, and usage guidelines.
-**Model weights are not publicly distributed.**
-Local access will be provided exclusively through the official JumpLander desktop/server application.
 ---
-## 📊 Current Status
-<img src="https://cdn-uploads.huggingface.co/production/uploads/69204763af796f2f22ad9f49/A8r0WUkLpEhDAh7Z8Xajx.jpeg" width="600"/>
-- ✔ Documentation available
-- ✔ Prototype benchmarks included
-- ✔ Online demo available on the website
-- ❌ Weights are *not* public
-- 🔒 Local model execution will be provided only via official software
 ---
-# 🧪 Benchmarks (Prototype)
-| Task | Score | Notes |
-|------|-------|--------|
-| **HumanEval** | **72%** | Strong execution accuracy |
-| **Repo-level Q&A** | High | Stable multi-file reasoning |
-| **Persian Instruction Following** | **Excellent** | Optimized bilingual performance |
----
-# 📦 Model Comparison (Prototype Benchmarks)
-| Model | Params | HumanEval | Multi-file Reasoning | Persian Support | Speed (tok/s) | Availability |
-|------|--------|-----------|------------------------|------------------|----------------|--------------|
-| **JumpLander Coder 32B** | 32B | **72%** | ✔ Strong | **Excellent** | 34 | Local-only via app |
-| Qwen2.5-Coder 32B | 32B | 75% | Medium | Weak | 32 | Open-source |
-| DeepSeek-Coder 33B | 33B | 79% | Strong | Weak | 29 | Open-source |
-| StarCoder2 15B | 15B | 63% | Limited | Weak | **45** | Open-source |
-| Llama-3.1 70B | 70B | **82%** | Strong | Weak | 20 | Open-source |
 ---
-# 💡 Why JumpLander Coder 32B?
-> ### 🧠 Multi-file reasoning
-> Designed for architecture-level understanding and full-repository analysis.
-> ### 🇮🇷 Persian-optimized workflow
-> Tuned for real Persian programming scenarios and instruction patterns.
-> ### 🛡️ Secure-by-design outputs
-> Refactoring logic, patch suggestions, and safe coding guidelines included.
-> ### ⚡ Developer-focused ecosystem
-> Future SDK, CLI tools, and integrated analysis modules.
 ---
-## 🗂 Local Execution (Official Software Only)
-Local execution of the model will be provided through the **JumpLander App**, enabling:
-- Secure local model loading
-- Offline and online operation modes
-- Integrated coding environment
-- Automatic model updates
-- Full repository understanding features
-**Note:**
-Weights will *not* be downloadable manually.
-They are packaged, encrypted, and tied to the official software.
 ---
-## 🎯 Use Cases
-- Application scaffolding
-- Repository-wide refactoring
-- Debugging & architecture inspection
-- Documentation and API specification
-- Programming education (EN + FA)
 ---
-## 🛠 Planned Capabilities
-- Repository-wide code generation
-- Multi-language support: Python, JS/TS, Go, Rust, Java, C/C++, Bash, SQL
-- Long-context reasoning (hundreds of thousands of tokens)
-- Test generation: unit, integration, regression
-- IDE extensions (VS Code + JetBrains)
-- Full SDK + CLI tools
----
-## 📎 Contact & Support
-Website: https://jumplander.org
-LinkedIn: https://www.linkedin.com/in/jump-lander-55812b388/
-Support: support@jumplander.org
 ---
-## 💻 Example Usage (Future API)
-```python
-from jumplander_sdk import JumplanderClient
-client = JumplanderClient(api_key="YOUR_KEY")
-# Scaffold a FastAPI app
-project = client.scaffold(
-    "Create a FastAPI service with JWT and PostgreSQL",
-    language="python"
-)
-project.save("./generated_app")
-# Refactor an existing repository
-patches = client.refactor("./myrepo", intent="improve structure")
-client.apply_patches(patches)

 ---
+# 🚀 JumpLander Coder 32B
+**Advanced Code‑Generation LLM — optimized for Persian‑speaking developers**
+**Short summary**
+JumpLander Coder 32B is a high‑performance, bilingual (English–Persian) code generation model optimized for multi‑file reasoning, repository‑scale analysis, and developer workflows. It is designed to assist with scaffolding, refactoring, testing, and documentation generation while emphasizing secure coding patterns and reproducible evaluation.
+> **Important:** Model weights are distributed **locally** through the JumpLander App (desktop/server installer). The model can also be tried on our website demo with limited free requests for evaluation. We do **not** publish model weights on an open public hosting by default — distribution is controlled via the official JumpLander software to ensure integrity and support.
+---
+## 🌟 Key Features
+- High‑quality, executable code generation and scaffolding
+- Multi‑file and architecture‑level reasoning
+- Secure‑by‑design outputs and automated refactoring suggestions
+- Persian (Farsi) instruction tuning for improved developer UX
+- CLI / SDK integrations and future IDE plugins planned
 ---
+## 📦 Local Distribution & How Users Access the Model
+JumpLander distributes model weights to end users via the official JumpLander App (installer) and controlled download endpoints. The purpose of local distribution is to enable offline and private execution, reduce API costs, and give users full runtime control on their machines.
+Typical flow (once local package is released):
+1. User installs JumpLander App (desktop or server).
+2. User downloads model bundle from the official server through the App (signed + checksummed).
+3. App verifies the integrity (SHA‑256 + PGP) and unpacks the model into a secure local runtime.
+4. The model runs locally — accessible via App UI, CLI, or local SDK.
+While the local installer is being finalized, a demo endpoint on the website provides limited testing (e.g., 100 trial requests) so users can evaluate model behavior without installing.
 ---
+## 🧪 Reproducible Evaluation & Benchmarks
+We publish reproducible evaluation scripts and raw logs so independent researchers can reproduce our reported numbers. Evaluation artifacts include:
+- `scripts/run_humaneval.py` (example)
+- `scripts/run_repo_reasoning.py`
+- Raw logs under `eval_logs/` with seeds and environment notes (CUDA/PyTorch versions)
+Example command (when you have a local model path):
+```bash
+python scripts/run_humaneval.py --model-path /path/to/jumplander-coder-32b --seed 42 --output eval_logs/humaneval.json
+```
+Metrics usually reported: pass@k (HumanEval), execution accuracy, latency (tokens/sec), and memory footprint.
 ---
+## 🔐 Integrity & Security (how downloads are verified)
+All published model bundles (when distributed) include:
+- `model.safetensors` (preferred safer serialization format)
+- `model.safetensors.sha256` (SHA‑256 checksum)
+- `model.safetensors.sig` (PGP detached signature)
+Example verification commands (Linux/macOS):
+```bash
+# Verify checksum
+sha256sum -c model.safetensors.sha256
+# Verify PGP signature (requires maintainers' public key)
+gpg --verify model.safetensors.sig model.safetensors
+```
+A convenience script `verify.sh` is included in this repository to automate the checks before loading the model locally.
 ---
+## 🛠 Quick example (Local Python loader)
+This example assumes the model files are verified and stored locally. The official App exposes a runtime; this snippet demonstrates the local loader pattern (trusted code only):
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+tokenizer = AutoTokenizer.from_pretrained("/local/models/jumplander-coder-32b")
+model = AutoModelForCausalLM.from_pretrained(
+    "/local/models/jumplander-coder-32b",
+    trust_remote_code=False  # We avoid remote code execution by design
+)
+prompt = "Create a simple FastAPI server with a single endpoint that returns 'hello'."
+inputs = tokenizer(prompt, return_tensors="pt")
+outputs = model.generate(**inputs, max_new_tokens=256)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
 ---
+## ✅ Trust & Transparency — Practical steps we follow
+To increase trust and demonstrate non‑fraudulent operation, JumpLander follows these practices:
+- Official distribution only through JumpLander App and controlled download endpoints.
+- Model bundles published with SHA‑256 checksums and PGP signatures.
+- Reproducible benchmarks and raw logs published in `eval_logs/`.
+- Public team profiles and contact information for accountability.
+- A demo endpoint (limited free requests) so users can validate model behavior before download.
+- Security guidance: run models in isolated environments, avoid `trust_remote_code=True` unless code is reviewed and signed.
+These steps are what we recommend including on the project page and in the model card to reassure enterprise and technical users.
 ---
+## 📁 Repository layout (suggested)
+```
+jumplander-coder-32b/
+├─ README.md
+├─ LICENSE
+├─ models/                    # (populated when bundles are released)
+│  ├─ model.safetensors
+│  ├─ model.safetensors.sha256
+│  └─ model.safetensors.sig
+├─ scripts/
+│  ├─ verify.sh
+│  ├─ run_humaneval.py
+│  └─ run_repo_reasoning.py
+├─ eval_logs/
+└─ docs/
+```
 ---
+## 📝 Contact & Support
+JumpLander Team — https://jumplander.org
+Support: support@jumplander.org
+LinkedIn: https://www.linkedin.com/company/jumplander
+---
+## Short Persian note
+🇮🇷 **جامپلندر — تجربهٔ توسعه برای فارسی‌زبانان.**
+در حال حاضر می‌توانید مدل را از طریق دموی وب سایت امتحان کنید؛ نسخهٔ محلی و نصب از طریق نرم‌افزار JumpLander عرضه خواهد شد.
+برای پشتیبانی و گزارش مشکلات، لطفاً به support@jumplander.org ایمیل بزنید.
+---