I study what post-training actually changes in small open-weight language models.
Two frozen studies on small open-weight models, and a retrieval tool built to keep research sources checkable. Each result below is shown with the number that complicates it.
Can post-training teach a 2B model when to call a tool, and what does it break?
0.6264→0.7548
call F1, 1,277-example pre-registered internal partition
The same model refused all 1,319 zero-shot GSM8K questions and lost 14.9 points 8-shot.
How much of a small model’s memory gap can retrieval and LoRA post-training close?
13.75%→71.25%
abstention on 80 unanswerable probes (of 688)
Answer accuracy rose too, but only to 18.59%.
Can a research tool return sources with locators, and say “unknown” instead of guessing?
0
fabricated answers on a frozen 40-task real-paper holdout
Fact recall 0.971 on the same holdout, but fact precision only 0.62.
I help build a nonprofit research effort focused on the parts of machine learning that are easiest to hand-wave: hardware, training, data, and the measurements that connect them.
Post-training, distillation, and evaluation of small open-weight models. Pre-registered where it counts, with the regressions and negative results left in.
Full-parameter SFT then DPO on Qwen3.5-2B lifted call F1 from 0.6264 to 0.7548 on a 1,277-example pre-registered internal partition and cut over-calling from 62.4% to 15.3%. The promoted model also refused all 1,319 zero-shot GSM8K questions and lost 14.9 points on 8-shot GSM8K, neither of which the promotion gate could see. One seed, reported with the regression left in.
Retrieval plus LoRA SFT, DPO and on-policy distillation took gemma-4-E2B from 0.16% to 18.59% judge-scored accuracy and from 13.75% to 71.25% abstention on unanswerable probes, over a 688-probe benchmark built to catch false memories. The abstention gain is large; absolute recall stays low. Single seed, with the errata published.
The hybrid system beats both a tuned deterministic pre-pass and the model alone, on every metric, on every checkpoint tested — field F1 0.6432–0.6827 against 0.2911 for rules alone and 0.4461 for the model alone. It also has a dominant, unsolved failure: silent omission, 55–62% of every failure breakdown, immune to four independent fixes. And a non-deterministic teacher labeller quietly invalidated part of the project's own early evidence before an audit caught it.
V3 targets the one number V2 could not move: omission at 55–62% of failures. Four iterations in, the best configuration reaches 52.3% — a new low, and still not below half. No V3 checkpoint beats the V2 release candidate on the comparable eval. This is an interim account of work in progress, not a result.
The engineering the research runs on: evaluation harnesses, retrieval and memory systems, and the products built alongside.
A schema registry, a deterministic pre-pass, and a distilled ~1B model routed by declared field ownership — beating both halves alone on every metric. 97 tests, a teacher-validation gate that rejects two in five labels before they can become training data, and every invariant enforced in code, not just in a docstring.
Five domain skills for MoE engineering — architecture, training, debugging, performance, benchmarking — packaged as a runtime-agnostic spec any frontmatter-reading agent can load, plus the CLI that deploys them and the four-arm harness built to test them against a search-enabled control instead of an information-starved strawman.
A personalized companion on gemma-4-E2B with a tiered, per-persona external memory: hybrid BM25 + dense retrieval fused by RRF, and LoRA SFT, DPO and on-policy distillation on top. On a 688-probe memory benchmark, abstention on unanswerable questions rose from 13.75% to 71.25% while accuracy stayed under 20%. 504 tests, a 112-entry claims registry, and 34 published errata.
Support, not endorsement — none of these organizations review or approve what I publish.
Competitions and recognition, judged against everyone else in the room.
Vendor certifications, each one linking to its own verification page.
Arjhine Ty
ML Research Engineer
Model
