Saheb Motiani

False self-correction in Llama-3.1-8B-Instruct

Writeup · Code and experiments

• Llama rejected correct arithmetic answers for some numbers (e.g. 100, 404) but not others. On 22 cases where it rejected at baseline, three MLP activation edits cut rejection to 3/22.
• Tested the edits with held-out unit selection and random controls; reverse patches also made selected, previously non-rejecting prompts produce rejection language.
• Tests of these two numbers suggested different early MLP components and a shared later influence on correctness judgments.

RL fine-tuning OLMo to withhold prohibited exam help

Writeup · Code and experiments

• Built a custom PyTorch/Transformers/PEFT GRPO trainer for OLMo-3-7B-Think-SFT with LoRA, vLLM rollouts and automated rewards; compared final-answer and whole-completion loss over 12 updates and 864 rollouts per run.
• Final-answer-only training increased rewarded answer tokens' relative likelihood in 47/51 eligible groups when saved reasoning was provided, yet fresh responses did not clearly improve.
• Whole-completion loss raised withholding from 68/144 to 81/144 fresh responses to seen prompts. Without policy text, it never refused prohibited help or asked permission (0/24 each on development prompts), so gains remained policy-dependent. One run per loss scope; automated judgments.

• Built notifications prompting non-admin users at high-value advertisers to appeal mistaken ad-account enforcements, unblocking $6.6M/year in revenue.
• Ran experiments showing broader appeal notifications increased harm, as bad actors exploited gaps in the appeal process. Worked with data scientists and the abuse-prevention team to limit the rollout to high-value advertisers.
• Fixed a multi-recipient concurrency bug, reducing notification failure rates by 70% for Facebook Pages, 30% for ad accounts and 8% for businesses; found the platform logging false-positive delivery successes.

• Designed a Python/Faust pattern for deploying ML risk-prediction models, with outputs consumed by clinical portals and patient-outreach tools.
• Built reusable Flink patterns for ETL, feature engineering and stream processing; the data platform team began rolling out Flink internally.
• Led an 8–10-person cross-functional team delivering a health-scoring product with Applied Scientists; developed Scala services for health-assessment and symptom-checker chatbot workflows.

• Designed distributed services to validate incoming media before transcoding and packaging, connecting on-premises infrastructure and AWS through asynchronous Kafka communication.

Orbit Wars – self-play RL training and evaluation system

Kaggle placement: 1,060 / 4,729 · Code and results · Interactive replay

• Built a GPU-vectorized PyTorch simulator, entity-transformer PPO policy, self-play league and standalone export; measured ~4.1k end-to-end environment steps/s on RTX PRO 6000, including policy inference and PPO updates.
• Diagnosed an evaluator that counted opponent crashes as wins; added fail-closed checks, built fixed, seat-balanced evaluation panels and checked train/export parity.
• A post-competition policy won 98.0% against a mid-tier planner and 14.1% against the 13th-ranked agent (256 games each), but 0/32 against a top-10 agent.

Snake – board-size generalization

Study and code · Demo

• An attention-pooling agent with egocentric inputs, trained on 10×10 and 20×20 boards, achieved 97.7% moves toward food in zero-shot 100×100 play. Probes and training-size ablations pointed to training-state coverage as a bottleneck.