Skip to content
View shipbehaves's full-sized avatar
2 followers · 7 following

Achievements

Achievement: YOLO

Achievements

Achievement: YOLO

Block or report shipbehaves

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
shipbehaves/README.md

Emre Yavuz · @shipbehaves

ships ai in places where it has to behave.

post-training across the stack. rlhf, rlvr, constitutional ai, and a 2.5b turkish lm pretrained from scratch. focused in rl and reward design, hands on with datasets.

yavuz.ai · shipbehaves.github.io · huggingface.co/yavuz-ai · x.com/yavuzai

i train, break, and measure language models and their reward signals, and write down what actually happened, nulls included. write-ups and results live on the research hub.

selected work

a few representative pieces.

  • pretrain · oghuz. a 2.5b turkish lm pretrained from scratch and post-trained to decline what it cannot verify. refusal generalized by sentence shape, not intent. teaching the intent across eight framings took adversarial refusal from 30% to 73% with helpfulness up. evaluation gallery, live demo, and tokenizer are public.
  • train · grpo-gsm8k. grpo + rlvr on gsm8k, no reward or value model. about 72s per step without vllm rollouts, about 2s with vllm colocate, roughly 30x. the algorithm was never the bottleneck, the rollout was.
  • break · self-reward-collapse. does a model training on its own judgment collapse? on verifiable math, not really. a brevity signal gets reward-hacked and halves answer length, but capability holds. the honest failure was more interesting than the headline would have been.
  • measure · regulated-evals. regulation-anchored trustworthy-ai scorecards for frontier and open-weight models in regulated finance. ten models in the september 2026 run, claude opus 5 and sonnet 5 included. none clears the bar on its own. every finding carries a severity, its binding-law citation, and an evidence-confidence tier, and every verdict reproduces from frozen transcripts. scorecards →
  • lifecycle · ai-delivery-contract. an industry-neutral reference for taking ai from data to compliant production, with a runnable governance-as-code gate. the model is one component in the system. the contract around it is what gets it shipped. interactive walkthrough →

open-source

models + datasets

huggingface.co/yavuz-ai · trained adapters, reward models, the preference and trajectory datasets from the work above, and the oghuz tokenizer.

Pinned Loading

  1. constitutional-cai constitutional-cai Public

    Constitutional AI reproduction (Bai et al. 2022) on a small open model: self-critique/revise SFT + RLAIF DPO, two-axis safety/over-refusal eval, and a failure analysis of the over-refusal regression.

    Python

  2. distributed-sft-fsdp distributed-sft-fsdp Public

    Genuine multi-GPU FSDP full fine-tune of a 7B across 4x A100 (closes the scale gap), with the tied-embeddings, collective-save, and checkpoint-consolidation gotchas made concrete.

    Python

  3. grpo-gsm8k grpo-gsm8k Public

    GRPO + RLVR on GSM8K (DeepSeek-R1 / TinyZero recipe): verifiable reward, no reward/value model, with the vLLM-rollout necessity and headroom lessons made concrete.

    Python

  4. regulated-evals regulated-evals Public

    Reproducible, regulation-anchored Trustworthy-AI scorecards for frontier and open-weight models in regulated industries (finance first).

    Python

  5. reward-model-ppo reward-model-ppo Public

    Classic RLHF: train a reward model (0.757 held-out) then PPO a policy against it, with an honest teardown of PPO's memory + instability cost vs DPO.

    Python

  6. self-reward-collapse self-reward-collapse Public

    Does a model training on its own judgment collapse? On verifiable math: the reward gets hacked (a brevity reward halves answer length) but capability does not collapse. An honest failure analysis w…

    Python