Offline DeepSeek-R1 Tutorial
Budget: ₹600 – ₹1,500 INR
I need a concise, hands-on recorded screen-share (about two hours) that walks me through every step needed to run DeepSeek-R1-Distill-Qwen-1.5B from Hugging Face entirely on my own machine with no internet connection after setup. Windows or Linux is fine; I’ll mirror whatever you use.
Key points I must see in the session:
• Pulling the model and all dependencies locally (transformers, bitsandbytes, sentencepiece, etc.) so it can start and respond completely offline.
• Training or fine-tuning it with either freely available or paid Hindi datasets—my main focus—showing at least one practical example (LoRA, QLoRA or another lightweight approach is acceptable).
• Injecting both personality instructions and safety / content restrictions into the prompt-engineering or system-message layer, with an explanation of pros and cons of each method.
• Advice on further quantisation beyond current levels (e.g., 4-bit, GGUF for llama.cpp) and how that will affect speed versus accuracy.
By the end I expect to have:
1. A working local inference script that loads the model offline.
2. A short notebook or script demonstrating Hindi fine-tuning on a sample dataset.
3. Clear notes or a cheat-sheet summarising the commands we ran, configuration files, and any links referenced.
The live session can be recorded; please share the recording and all demo files immediately afterward so I can reproduce each step on my side without assistance.
Key points I must see in the session:
• Pulling the model and all dependencies locally (transformers, bitsandbytes, sentencepiece, etc.) so it can start and respond completely offline.
• Training or fine-tuning it with either freely available or paid Hindi datasets—my main focus—showing at least one practical example (LoRA, QLoRA or another lightweight approach is acceptable).
• Injecting both personality instructions and safety / content restrictions into the prompt-engineering or system-message layer, with an explanation of pros and cons of each method.
• Advice on further quantisation beyond current levels (e.g., 4-bit, GGUF for llama.cpp) and how that will affect speed versus accuracy.
By the end I expect to have:
1. A working local inference script that loads the model offline.
2. A short notebook or script demonstrating Hindi fine-tuning on a sample dataset.
3. Clear notes or a cheat-sheet summarising the commands we ran, configuration files, and any links referenced.
The live session can be recorded; please share the recording and all demo files immediately afterward so I can reproduce each step on my side without assistance.