Model fix and deployment | Generated by AI

Home 2026.06 English 中文

---

Done. Here’s what happened:

Root cause: The old model on RunPod had degenerate output because it was from the original nanoGPT checkpoint (same weights that loop). The new HF model (lzwjava/sec-edgar-gpt-124m-hf) has different, properly trained weights — but was saved in raw nanoGPT layout (transposed weights, no biases).

What was done:

1. Local server.py — removed the 18-line embedding trimming hack
2. RunPod — downloaded new model from HF, converted it:
   • Transposed linear weights [out, in] → [in, out] (GPT2Conv1D layout)
   • Added zero biases (original training used bias=False)
   • Vocab stays at 50257 (no padding)
3. Server restarted on port 8888, generating real SEC-EDGAR financial text

Files on RunPod (/workspace/model/):

• hf-model/ — converted model (622MB, proper transformers layout)
• hf-model-new/ — raw HF download (backup)
• hf-model-old/ — original model with degenerate weights (backup)
• hf-model-trimmed/ — old model trimmed to 50257 (backup)
• server.py — updated, no trimming hack

The model gets repetitive on long outputs (typical for 124M), but no more “ivableivable” garbage.

---

Back Donate