Foundation Model Systems · LLM Post-training · ML Research Engineering
Building and evaluating model systems from first principles — with reproducible experiments, measured trade-offs, and honest failure analysis.
B.Sc. Information and Computing Science · Hefei University of Technology · 2026
| Track | Current focus | Public status |
|---|---|---|
| Foundation model systems | Tokenizer → Transformer → training & systems | WendyFM: tokenizer milestone completed |
| Post-training research | Verifier reliability, mathematical difficulty, SFT-to-RL questions | Ongoing · research repository private |
| Applied machine learning | Time-aware evaluation, retrieval, ranking, ablation | Offline recommendation reproduction |
Shipped milestone: a deterministic byte-level BPE tokenizer with two trainers, multilingual correctness tests, and CPU-local benchmarks.
| English | Code | Mixed multilingual | Chinese |
|---|---|---|---|
| 7.25× faster | 4.55× faster | 4.57× faster | 0.70× (slower) |
Indexed trainer vs. naive reference on small synthetic corpora. All specifications matched exactly; indexed training had higher tracked Python allocation peaks. These figures are not LLM pretraining, GPU, or production results.
Code · Benchmark report · Measurements (CSV) · Reproduction steps
Next (not yet completed): Decoder-only Transformer engineering draft — PR #24 has passed isolated toy CPU smoke tests and awaits owner review, environment reproducibility checks, and real training/evaluation. GPU/attention profiling and scaling remain planned.
Exploring how data selection, difficulty profiling, and answer verification could influence learning under a fixed compute budget. Current work emphasizes verifier agreement, conservative extraction, and experimental design; SFT/RL findings and publication are not claimed.
Research code stays private until validation and release review. This section will be updated only when safe, reproducible findings are ready.
A public offline reproduction of a Two-Tower/DSSM short-video recall pipeline, with temporal splits, negative sampling, Faiss-style TopK retrieval, ablations, leakage checks, and failure analysis.
Other public evidence: multimodal retrieval · RAG retrieval evaluation · Transformer model lab.
Question → baseline → implementation → tests → measured comparison → limitations.
I prefer mature, properly credited open-source foundations over unnecessary reinvention. I care about correctness, negative results, clear reproduction commands, and practical engineering constraints — not the number of green squares on a contribution chart.
Industry experience informs my engineering approach, but employer-owned data, internal model artifacts, implementation details, and private metrics are not portfolio content. Public sample-based experiments are not presented as online business outcomes.
Responsible publication policy · Interested in discussing foundation models, training systems, reproducible evaluation, or research engineering? Get in touch →