Today's FinGPT Training Plan | Generated by AI

Home 2026.06

Based on your notes from the last 24 hours, here’s what I see you’ve been deep-diving on and what makes sense to do today:

Your Recent Focus (Last 24h)

  1. Financial LLMs — BloombergGPT vs FinGPT deep dive
  2. LoRA fine-tuning — contrarian truths, practical limits
  3. RTX 4070 compute budget — what you can actually train (300-500M params)
  4. Blog infrastructure — Jekyll/GitHub Pages migration constraints
  5. SGLang/Triton — getting inference working on your workstation
  6. CC BY 4.0 licensing — for your notes knowledge base

What To Do Today

1. Ship the FinGPT-style experiment on your RTX 4070 (Highest Priority)

Your notes already laid out the exact recipe — take Qwen or Llama-3-8B, build a small financial sentiment instruction set from headlines, QLoRA-tune on your 4070. This is the portfolio piece that bridges your banking domain + AI engineering transition. You already have SGLang serving working, so you can test immediately after training.

# Start on your workstation
ssh lzw@192.168.1.36
# Clone FinGPT, adapt their pipeline to your hardware

2. Fix the SGLang/Triton segfault (Unblock Inference)

You hit a Triton 3.5/3.6 segfault on Ada architecture. Your notes say Triton 3.3.1 + cu126 avoids it. Make sure ~/sglang-env is stable so you can serve models after fine-tuning.

3. Register lzwjava.com ($10/yr)

Your notes concluded: owning the apex domain is worth it for portability. You have 400+ posts and 8,000 notes — being locked to *.github.io is unnecessary risk. This is a 10-minute task.

4. Apply CC BY 4.0 to your notes

You researched the license. Now apply it — add a LICENSE file to your notes repo with CC BY 4.0. Quick win.

5. Set up rsync backup to Mini M4

Your notes mention rsync 2.6.9 drops on big transfers but re-run resumes. Run mini-backup.sh to huarong@192.168.1.54 — get your 25GB backup current.

The Bigger Picture

Your notes show you’re building the exact stack that matters: training (nanoGPT/LoRA), serving (SGLang), domain expertise (financial LLMs), and infrastructure (GPU machines, backups). The FinGPT experiment is the highest-leverage thing because it produces a tangible artifact that demonstrates all four.

What do you want to tackle first?

Back Donate