Describe your research and contributions

# How to deliver a high-impact 10–15 minute research overview (and an example answer) ## A. Recommended structure and time budget - 0:00–1:00 — One-sentence summary and why it matters - 1:00–2:30 — Problem and prior work - 2:30–4:30 — Novel contributions (3 bullets max) - 4:30–7:00 — Methodology (diagram-worthy; equations only if clarifying) - 7:00–9:30 — Experimental setup (datasets, metrics, baselines, hardware) - 9:30–12:00 — Results and ablations (numbers, not prose) - 12:00–13:30 — Limitations and failure cases - 13:30–15:00 — Your impact, collaboration/timeline, and next steps Tip: Keep a running thread of the causal story: the choices you made → the evidence → the impact. ## B. Slide outline (6–8 slides) 1) Problem and why it matters 2) Related work and gap 3) Your contributions (bulleted) 4) Method overview (diagram + key equations) 5) Experimental setup (datasets/metrics/baselines/hardware) 6) Results (main table or bullet list of metrics) 7) Ablations and sensitivity 8) Limitations, your impact, timeline, and next steps ## C. Worked example talk track (adapt to your project) Below is a realistic example in the LLM efficiency/instruction-tuning space. Replace with your own project but mirror the structure and level of specificity. ### 1) Problem and motivation - Problem: Achieve strong instruction-following performance with a 7B–13B LLM using far less compute than typical full fine-tuning, while improving evaluation scores and reducing toxicity. - Why it matters: Smaller models with strong chat performance reduce serving cost and latency, enable on-device/edge use, and broaden access. One-liner: We developed a data-driven instruction-tuning and preference-optimization pipeline that matches or exceeds common 13B chat baselines using a 7B–13B model with ~40% less fine-tuning compute. ### 2) Related prior work (gap) - Self-Instruct, FLAN, Vicuna: Show instruction tuning works, but sensitive to data quality and can overfit style. - RLHF with PPO: Strong results but compute- and infra-heavy. - DPO/KTO: Preference learning without online RL; simpler to train but sensitive to reference choice and hyperparameters. - LIMA: Few, high-quality examples can go far; highlights the importance of data quality over volume. Gap: Prior work underexplores systematic data dedup/quality filtering and length-aware preference optimization at small scales with rigorous, multi-metric evaluation. ### 3) Novel contributions - Data quality and balance at scale: MinHash deduplication, perplexity filtering, and domain-balanced sampling across general, reasoning, and safety data. - Length-normalized Direct Preference Optimization (DPO): A simple modification that stabilizes learning and reduces verbosity bias. - Lightweight systems optimizations: FlashAttention2, FSDP Stage 3, gradient checkpointing, mixed precision, and token-based scheduling for stable throughput. ### 4) Methodology - Base model: Open 7B and 13B dense decoders with RoPE positional embeddings. - Supervised fine-tuning (SFT) on curated instruction data to initialize an aligned policy. - Preference learning with DPO: - For triplets (x, y+, y−), minimize the loss L_DPO = − E[ log σ(β(Δlogπθ − Δlogπ_ref)) ] where Δlogπθ = log πθ(y+|x) − log πθ(y−|x), and Δlogπ_ref uses a frozen reference model. - Length-normalization: scale Δlogπ terms by 1/len(y) to counter verbosity bias. - β tuned via ablations (0.2 worked best here). - Systems details: FSDP3 + FlashAttention2 + fused ops; bfloat16; gradient clipping 1.0; cosine LR schedule with 3% warmup; global seed control. Compute sanity check: Effective tokens = sequence_length × global_batch_size × steps. Example: 2048 × 256 × 6000 ≈ 3.1B tokens. ### 5) Experimental setup - Datasets - SFT: ~1.2M instruction–response pairs (FLAN v2, OpenOrca, UltraChat, curated StackExchange explanations), deduped with MinHash; perplexity-filtered using a small LM to remove low-signal text; domain-balanced ratio 50% general, 30% reasoning, 20% safety. - Preference: 50k pairwise comparisons (public helpful/harmless and curated synthetic pairs). Held-out 5k for validation. - Metrics and evaluation - Instruction-following: MT-Bench (8 domains, 2-turn), AlpacaEval 2 pairwise win rate. - Knowledge/reasoning: MMLU (5-shot), HellaSwag (0-shot), GSM8K (8-shot-CoT), TruthfulQA. - Safety/toxicity: Perspective API toxicity rate; harmlessness on HH-RLHF subset. - Calibration: Length-normalized log-prob and response length distribution. - Baselines - 7B SFT-only, 7B DPO (naïve), 13B SFT-only, 13B chat model from public checkpoints; Zephyr-7B as a strong community baseline. - Hardware/compute - Training on 64× A100 80GB, FSDP3, FlashAttention2; token throughput ~180k–220k tok/s. - Fine-tuning compute reduced by ~40% vs full SFT+DPO baseline via LoRA for SFT and efficient packing. - Reproducibility - 3 seeds for key results; bootstrap CIs for win rates; strict train/val/test splits and dedup against eval sets. ### 6) Results (illustrative but realistic) - Instruction-following - MT-Bench: 7B SFT 6.4 → Ours (7B SFT+DPO+LN) 7.6; 13B SFT 7.1 → Ours (13B) 7.9. - AlpacaEval 2 win rate vs 13B SFT baseline: 7B ours 62% (±2.5), 13B ours 68% (±2.1). - Knowledge/reasoning - MMLU (5-shot): 7B SFT 46.0 → 51.8; 13B SFT 53.2 → 56.1. - HellaSwag: essentially unchanged (7B ~76.5% → 76.8%). - GSM8K (8-shot-CoT): 7B 24% → 31%; 13B 34% → 39%. - Safety/toxicity - Toxicity rate reduced from 4.3% → 2.7% on safety probes with data mixture + DPO. - Efficiency - Fine-tune wall-clock reduced ~35–45% at similar or better scores via LoRA for SFT, length-normalized DPO, and packing. Takeaway: With targeted data curation and length-aware DPO, a 7B model matches or beats common 13B SFT baselines on instruction metrics, narrowing the gap on knowledge/reasoning, while reducing compute. ### 7) Ablations and sensitivity - Data deduplication: Removing MinHash dedup dropped MT-Bench by −0.8 and increased toxicity by +0.6 pp. - Domain mix: Shifting reasoning fraction from 30% to 15% cut GSM8K by −3.2 points with minimal MT-Bench change. - DPO β parameter: β = 0.2 best; β = 0.05 underfits; β = 0.5 over-penalizes and increases verbosity. - Length normalization: Removing it raised average response length +18% and reduced AlpacaEval win rate by −3.1 pp. - LoRA vs full SFT: LoRA recovered ~98% of SFT performance at ~18% of compute for SFT stage. - Seeds: Variance of ±0.2–0.4 MT-Bench; report means and CIs. ### 8) Limitations and failure cases - Math and multi-step reasoning still lag specialist models; CoT helps but is brittle. - Long-context (≥16k) support not addressed; partial degradation past 8k without RoPE rescaling. - Style over-optimization: Some verbosity or hedging remains on safety prompts. - Sensitivity to preference data noise; poor pair quality can destabilize DPO. Failure examples - Multi-hop QA with distractors: Generates plausible but incorrect intermediate steps. - Safety edge cases: Over-refusal on benign biomedical queries. ### 9) Individual impact and ownership - Led methodology: implemented length-normalized DPO, reference-policy smoothing, and evaluation harness (∼3.5k LOC across trainer, metrics, and scripts). - Designed data pipeline: MinHash dedup, perplexity filter, and domain mixing; owned the ablation plan. - Ran 50+ training/eval jobs; created dashboards for MT-Bench, MMLU, and toxicity with CIs. - Made systems optimizations (packing, fused kernels, FlashAttention2) yielding +1.7× training throughput. Quantified impact - Delivered +1.2 to +1.5 MT-Bench over 7B baselines; −35–45% fine-tuning compute; −1.6 pp toxicity. - Authored the internal report and reproducible configs; unblocked integration into serving. ### 10) Collaboration and timeline - Team: 1 research engineer (me), 1 data engineer, 1 infra engineer, 2 part-time annotators. - Timeline (12 weeks) - Weeks 1–2: Baseline reproduction (7B/13B SFT), eval harness. - Weeks 3–4: Data pipeline (dedup, filtering, mixing) + sanity checks. - Weeks 5–7: DPO variants, β sweep, length normalization. - Weeks 8–10: Ablations (data mix, seeds, LoRA vs full SFT), safety eval. - Weeks 11–12: Hardening, docs, handoff to serving. ### 11) What I would do differently and next steps - Differently: Start with a preregistered ablation plan and automated seed sweeps; add early human eval to reduce reward hacking. - Next steps - Preference data quality: Self-consistency filtering and annotator calibration. - Long-context: RoPE rescaling/YARN; curriculum to 32k tokens. - Reasoning: Lightweight tool-use or synthetic scratchpads; selective CoT. - Safety: Multi-objective preference modeling to reduce over-refusals. ## D. Guardrails, pitfalls, and validation - Guardrails - Prevent data leakage: hash-based dedup against eval sets. - Track compute: tokens, FLOPs, wall-clock; enable apples-to-apples comparisons. - Report uncertainty: multiple seeds or bootstrapped CIs. - Pitfalls - Over-tuning on leaderboard sets; noisy preference data; comparing to weak baselines. - Confounding changes (e.g., both data and model changed) without ablations. - Validation checklist - Predefine metrics and stopping criteria. - Keep a single change per experiment when possible. - Log all settings; ensure deterministic seeds where possible. ## E. Template you can reuse - Title: One-line problem, one-line impact. - Problem and why it matters: Who benefits; what metric moves and by how much. - Prior work: 2–4 references and the gap. - Contributions: 3 bullets max. - Method: Diagram + 1–2 equations; training specifics. - Setup: Datasets, metrics, baselines, hardware; reproducibility notes. - Results: Numbered bullets with deltas; at least one ablation. - Limitations/failures: 2–3 concrete examples. - Your impact: Ownership percentage, key decisions, LOC, experiments run. - Timeline and collaboration: Roles, milestones. - Do next: 2–3 focused, testable follow-ups. This structure ensures you address exactly what interviewers look for: a clear problem, defensible methodology, credible evidence, honest trade-offs, and your direct impact.