╔═══════════════════════════════════════╗
║ >_ loss.backward() ║
║ epoch 47/100 ... diverging ║
║ retrying with lr=1e-5 ║
╚═══════════════════════════════════════╝
Fine-tuning runs, distillation pipelines, and training experiments — logged honestly, including the ones that didn't work. Publishing checkpoints to HuggingFace @ SamY36.
| Run | Base model | Method | Outcome |
|---|---|---|---|
vibethinker-selfdistill-v1 |
Qwen2.5-Coder-3B | Multi-teacher distillation (4 Groq-hosted teachers → 242 traces) + SFT | ✅ SFT converged, val loss 1.776 — uploaded |
vibethinker-grpo |
same as above | GRPO on top of SFT checkpoint | ❌ Abandoned — reward log anomalies, not trustworthy |
coding-lora-qwen3-4b |
Qwen3-4B-Instruct-2507 | LoRA r=64, Q8_0 GGUF export, multi-lang + cybersecurity + web dev | 🔄 In progress — mid-debug on Kaggle dependency pinning (unsloth/trl/datasets/torchao vs torch 2.6.0+cu124) |
monster_coding.jsonl |
— | Dataset pipeline: per-source schema extraction, dedup, <think> block wrapping |
✅ Built, feeding sub-3B training runs |
More entries get added as runs finish — successful or not.
Compiler design
| Repo | Stack | What it is |
|---|---|---|
| Compiler_f | Python | Compiler implementation |
| Compiler_final | Python | Final compiler project codebase |
| Compiler | — | Compiler design coursework |
| Shop_ai | Python | AI agent for e-commerce/shopping |
| Ai_Convert | Typescript | Plain-English into a fully validated, machine-readable application specification in seconds. |
