Projects

Everything here is public. Models on Hugging Face, code on GitHub. Download counts are read from the API when this page is built, not typed in by hand.

Post-Training 101: an RL recipe for text-to-SQL, in progress

Taking Qwen2.5-Coder-7B from no training through SFT, DPO and GRPO on BIRD, one part at a time, with the maths behind each step. The environment is open on GitHub. Part 1: Understanding the data and the benchmark.

A 7B that ties a 1.6T model at text-to-SQL

Post-training finetuning-text-to-sql

This was the first study. Post-Training 101 rebuilds it step by step, in the open.

Qwen2.5-Coder-7B, post-trained and benchmarked on BIRD dev (1,534 questions). Accuracy means result accuracy: every generated query is executed against the real database and the result sets are compared. No fuzzy string matching, no exact-match SQL scoring, because both of those lie.

ModelParamsAccuracy
GLM 5.2744B63.0%
DeepSeek V4-Pro1.6T58.7%
This work, 7B + Best-of-N 7B 58.5%

A 7B ties a 1.6T model at roughly 0.4% of the parameters, and runs locally. GLM 5.2 still wins outright. This is a parameter-efficiency result, not a superiority one. The two things that moved the needle were distilling the teacher's reasoning rather than its SQL (+5.2pp, and it generalized to unseen databases) and execution-based majority voting at inference (+6.4pp, no retraining).

Open models & datasets

Local 5.8k downloads all time.

Artifact Type All time 30 days
qwen2.5-coder-7b-bird-cot Model · reasoning-distilled text-to-SQL 4k 106
…-GGUF Model · quantized for local inference 875 139
bird-cot-sft Dataset · CoT traces from Qwen3-Coder-480B 242 66
spider-dpo-1040 Dataset · 1,040 execution-verified preference pairs 719 41
LocalSQL Space · ask a question, get SQL, run it live

All time is cumulative; 30 days is a moving window that decays after a spike. Shown separately because quoting one as the other is the easiest way to be wrong by an order of magnitude.

Detonating dependencies before your agent installs them

Agents package-quarantine

Supply-chain attacks like the LiteLLM and ctx PyPI hijacks fire at install time, so static CVE scanners miss them entirely. This intercepts pip and npm, runs the package in a Docker honeypot seeded with canary credentials, and watches at the OS boundary with strace, because postinstall hooks spawn child processes that in-process monkey-patching never sees.

Research

Research Automated Highlight Generation using Deep Learning in Soccer Matches, IEEE ICAD 2026. Turns a full soccer match into highlights by sorting 10-second clips into events like goals, fouls and red cards.

Also

Learning repos: GRPO with the SQL executor as the reward function, plus its RL environment. MedSignal runs clinical intelligence entirely on-device. SplitMate has split $23K+ across 182+ receipts for 22+ people.