Production over paper. I build LLM post-training pipelines and the infrastructure they run on — a 27-method fine-tuning spectrum with receipts on every number, production-resident decision models, GPU fleet ops, and agent tooling.
| Platform | Handle | What lives there |
|---|---|---|
| 🧩 ModelScope | theGAI | 45 repos — full 27-method spectrum, eval datasets, MCP server, field tutorials（中文主场） |
| 🤗 Hugging Face | jjshin | EN storefront — 7 repos + 📌 curated collection |
| 🐙 GitHub | jjxyxsjr | Infra & DR mirrors (private), agent tooling, MCP servers |
- kev-9B-LoRA-ranker — alert-triage ranker: LM backbone + 256-d rank head, abstain as a first-class citizen. Production-resident: 90% agreement with human arbitration vs 10% for the rules it replaced (123 real alerts, 40 adjudicated).
- Qwen3.5-4B-ReFT-gsm8k — ReFT v5 hits GSM8K 0.876 (+1.5pt over base) after fixing a 320-token truncation that cost v4 −44.6pt. Same base, same task — only the extraction pipeline changed: a 20.5pt ARC gap between two adapter versions.
- laya-fullft-typed-decisions — decision-head full-FT + RLCD with proper scoring rules, hand-written pure-PyTorch (no trainer could express it). ECE halved on hard suites — calibration is what makes a decision model deployable.
- kev-triage-probe 🚨 — interactive demo of the triage severity distribution. Zero backend, play with it.
- ms-hub-insight — MCP server giving agents ModelScope hub insight: lineage trees, metadata verification, market search. Public API only, zero token.
- trihub — one CLI to audit & operate a portfolio across GitHub / Hugging Face / ModelScope: mirroring gaps, model-card hygiene, cross-platform links, one-command model mirroring. Ships a CLI + MCP server (read-only by default, writes behind a dry-run gate).
Negative results ship too — the cards document what was tried and didn't move the needle, because that's where the engineering judgment lives.
中文一句话：LLM 训练管线 × MLOps 基础设施双线——27 法微调谱系全收据、告警分诊模型生产常驻、GPU 集群与灾备自动化。