Building Training Frameworks

If you’re an AI engineer, understanding how LLMs are trained and aligned is essential for building high-performance, reliable AI systems. Most large language models follow a 3-step training procedure: Step 1: Pretraining → Goal: Learn general-purpose language representations. → Method: Self-supervised learning on massive unlabeled text corpora (e.g., next-token prediction). → Output: A pretrained LLM, rich in linguistic and factual knowledge but not grounded in human preferences. → Cost: Extremely high (billions of tokens, trillions of FLOPs). → Pretraining is still centralized within a few labs due to the scale required (e.g., Meta, Google DeepMind, OpenAI), but open-weight models like LLaMA 4, DeepSeek V3, and Qwen 3 are making this more accessible. Step 2: Finetuning (Two Common Approaches) → 2a: Full-Parameter Finetuning - Updates all weights of the pretrained model. - Requires significant GPU memory and compute. - Best for scenarios where the model needs deep adaptation to a new domain or task. - Used for: Instruction-following, multilingual adaptation, industry-specific models. - Cons: Expensive, storage-heavy. → 2b: Parameter-Efficient Finetuning (PEFT) - Only a small subset of parameters is added and updated (e.g., via LoRA, Adapters, or IA³). - Base model remains frozen. - Much cheaper, ideal for rapid iteration and deployment. - Multi-LoRA architectures (e.g., used in Fireworks AI, Hugging Face PEFT) allow hosting multiple finetuned adapters on the same base model, drastically reducing cost and latency for serving. Step 3: Alignment (Usually via RLHF) Pretrained and task-tuned models can still produce unsafe or incoherent outputs. Alignment ensures they follow human intent. Alignment via RLHF (Reinforcement Learning from Human Feedback) involves: → Step 1: Supervised Fine-Tuning (SFT) - Human labelers craft ideal responses to prompts. - Model is fine-tuned on this dataset to mimic helpful behavior. - Limitation: Costly and not scalable alone. → Step 2: Reward Modeling (RM) - Humans rank multiple model outputs per prompt. - A reward model is trained to predict human preferences. - This provides a scalable, learnable signal of what “good” looks like. → Step 3: Reinforcement Learning (e.g., PPO, DPO) - The LLM is trained using the reward model’s feedback. - Algorithms like Proximal Policy Optimization (PPO) or newer Direct Preference Optimization (DPO) are used to iteratively improve model behavior. - DPO is gaining popularity over PPO for being simpler and more stable without needing sampled trajectories. Key Takeaways: → Pretraining = general knowledge (expensive) → Finetuning = domain or task adaptation (customize cheaply via PEFT) → Alignment = make it safe, helpful, and human-aligned (still labor-intensive but improving) Save the visual reference, and follow me (Aishwarya Srinivasan) for more no-fluff AI insights ❤️ PS: Visual inspiration: Sebastian Raschka, PhD

Your sales team is optimizing for the wrong metric, and it's costing you millions Most sales leaders are obsessed with pipeline coverage ratios. "We need 3x coverage to hit our number." "Generate more top-of-funnel activity." "Increase prospecting activity by 40%." But coverage ratios are a vanity metric that's actually destroying your team's performance. Here's why this thinking is backwards Traditional logic is the same old… More opportunities = Higher probability of hitting quota Build massive pipeline = Insurance against deal slippage BUT in reality Bigger pipelines create cognitive overload for reps Too many opportunities = Poor qualification and deal management Reps spread thin across 50+ "opportunities" instead of focusing on 15 real ones The highest-performing sales teams I work with have completely flipped this Instead of maximizing pipeline size, they maximize pipeline quality. The Quality-First Framework looks like this 1) Ruthless Qualification Standards Only deals with documented business impact, defined evaluation processes, and accessible buying teams make it into the pipeline. 2) Rep Capacity Management Each rep can effectively manage 12-15 active opportunities. Anything beyond that diminishes focus and results. 3) Stage Velocity Tracking Measure how fast deals move through stages, not how many deals exist in each stage. 4) Elimination Before Generation Before adding new opportunities, eliminate stalled ones. Clean pipeline = clear thinking. The math is crazy Team A: 200 opportunities, 15% close rate = 30 deals Team B: 100 high-quality opportunities, 35% close rate = 35 deals Team B wins with half the pipeline stress. Your reps aren't struggling because they need more opportunities. They're struggling because they can't focus on the right ones. Share with a leader who needs to hear this ^^

🚀 Join us for an engaging episode of Notes to My (Legal) Self featuring the distinguished Harry Borovick, General Counsel at Luminance and lecturer at the University of Law, London. 🌟 Harry recently contributed to the pivotal discussion on Legal Operations in the Age of Data and AI, blending academic and practical expertise. In this episode, Harry explores: AI in Legal Practice: How AI is transforming the legal field. Legal Education Evolution: The necessity of integrating AI skills into legal training. Balancing Skills: Balancing traditional legal skills with emerging tech demands. Meet Harry Borovick Harry shares his unique experiences from his role at Luminance, an AI company leading legal technology innovation, and contrasts it with traditional General Counsel positions. His insights provide a comprehensive look at the future of legal practice. 🎙️ Highlights: The transformation of legal practice through AI. The importance of integrating AI into legal education. Balancing traditional legal skills with technological advancements. The role of feedback loops in developing and using AI products. Preparing law students for the complexities of modern legal practice. 💡 Learning Outcomes: AI Revolution: Understand how AI is revolutionizing legal practice and education. Skill Balance: Learn the critical balance between traditional legal skills and emerging tech demands. AI Integration: Discover the significance of integrating AI training in law schools. Adaptation Strategies: Strategies for adapting to technological changes in the legal profession. Continuous Learning: The importance of continuous learning and personal ownership in legal careers. 🔗 Don’t miss this thought-provoking discussion! Listen via the link in the comments. ❓ Questions: Do you think AI skills should be a compulsory part of legal education? How can traditional legal skills be preserved amidst the rise of AI? What impact will AI have on the legal profession in the next decade? #LegalTech #AI #LegalEducation #ContinuousLearning #LegalPractice -------- 💥 I am Olga V. Mack 🔺 AI & transformative tech expert in product counseling 🔺 Educating & upskilling human capital for digital transformation 🔺 Championing change management in legal innovation & legal operations 🔺 Keynotes on the intersection of business, law, & tech 🔝 Connect with me 🔝 Subscribe to Notes to My (Legal) Self newsletter

Some of my hardest lessons as a sales leader came when figuring out how to setup and run training (learn from my mistakes!) Me as a new leader: "Great we have 10 topics we want to cover... let's do 1 a week. 2.5 months later we will have covered SO much ground!" 🙃 Training was more of a "box checking" exercise. Someone shared feedback on what they wanted to learn, and it got added to the list Having one 30 or 60 minute training on any topic is never sufficient, and I did the team a disservice So what was missing? And what did I seek to add later? 👉 Focus Instead of 10 topics, we might go into a quarter with 1-2 priority focus areas. The deeper engagement on a narrower topic is not unlike narrowing your focus on a smaller set of ICP accounts This creates room for practice, follow up sessions, different voices delivering the material, and ultimately makes the content stickier 👉 Engagement from other departments Where applicable, involvement from other departments can add incredible value to your training program. For instance, when you are training on a new product category, it is valuable to: - Hear firsthand from Product how it's built - Align your training timeline with Product Marketing so that materials are ready to go as the training commences - Work with Marketing so that messaging aligns to how you can sell it and everyone has the same talking points from day 1 - Work with Rev Ops to identify a market opportunity to apply your learnings - Have Sales Enablement help prepare uses cases in your sales tech stack 👉 A system to encourage accountability Once the trainings are delivered, how do you know that the sales team was paying attention? That can take many forms: - Group activity like pitch practice - Measuring adoption through tools like Gong - Contest/SPIF to encourage initial matching sales activity - Knowledge tests in your LMS (my least favorite) 👉 Repetition There's a reason Sesame Street used to repeat episodes during the week - once wasn't enough to get the message home! While your sales team isn't full of 3 year olds, similar principles apply Bottom line: instead of thinking about any topic as a single "training", think about creating "training programs" for your team 🎓 Tying it all together for a training on "New Product A" Week 1: Product & Product Marketing introduce the new offering Week 2: Outside expert/marketing/leadership deliver the industry POV Week 3: Team gets together to identify prospects and practice the pitch Week 4: Team provides feedback on material and prospecting plans are built incorporating the training Weeks 5-8: Measuring adoption through Gong. Shouting out strong adoption and privately helping laggards identify gaps in understanding Week 6: Short contest to encourage cross/up-sell opportunity creation Week 12: Revisit/Feedback #SalesEnablement #SalesTraining #LeadershipLessons #CorrCompetencies

It seems that even for companies like Meta, OpenAI, or Google, training an LLM from scratch with a very long context window, such as 128k to 1M tokens, isn't reasonable because: - The cost is prohibitively high. For instance, training LLaMA 2 with a 4k context window was approximately twice as expensive as training LLaMA 1 with a 2k context window (even though LLaMA 2 was trained on 40% more tokens). If this extrapolation holds, it would mean that extending the context window from 8k to 128k could result in at least a 16x increase in cost. - There are hardly any long-sequence datasets publicly available. In other words, you would need to create a dataset of sequences up to 128k or 1M tokens for the training to be meaningful. Currently, most people likely use synthetic data to stitch sequences together and somehow maintain coherence, but this approach is challenging and incredibly tedious. So, how do you produce long-context LLMs without breaking the bank? I suspected that for such “LongLLMs,” the pre-training was done in two stages. The first stage uses a lower context window with the majority of the dataset, and the second stage uses much less data but with an extended context window. Yesterday, I took the time to read the LLaMA 3 405B paper and came across this sentence: “We pre-train a model with 405B parameters on 15.6T tokens using a context window of 8K tokens. This standard pre-training stage is followed by a continued pre-training stage that increases the supported context window to 128K tokens.” So, there it is! Even though this approach works, I suspect the model may retain more "bias" towards the shorter sequences seen during the initial training, potentially affecting its ability to handle long sequences with the same accuracy, thus leaving room for hallucinations and loss of information. The question now is, how do you properly evaluate long-context LLMs? Such evaluations would obviously be quite costly and time-consuming. I wanted to perform some empirical tests to evaluate how accurate the LLM is for different sequence lengths (I'm expecting to see a decline in accuracy after 20-30k tokens) but couldn't find any benchmarks or evaluations that were satisfying. Any ideas?

Exciting breakthrough in long-context language models! Microsoft researchers have developed a novel bootstrapping approach that extends LLM context lengths to an impressive 1M tokens while maintaining strong performance. >> Key Innovation The team introduces a clever self-improving workflow that leverages a model's existing short-context capabilities to handle much longer contexts. Rather than relying on scarce natural long-form data, they synthesize diverse training examples through: 1. Instruction generation using short-context LLMs 2. Document retrieval with E5-mistral-7b 3. Recursive query-focused summarization 4. Response generation >> Technical Details Their SelfLong-8B-1M model achieves remarkable results: - Near-perfect performance on needle-in-haystack tasks at 1M tokens - Superior scores on the RULER benchmark compared to other open-source models - Progressive training strategy with RoPE base frequency quadrupling at each stage - Efficient training using RingAttention for distributed processing - Implementation of PoSE-style training for hardware constraints - Utilizes vLLM for inference optimization >> Impact This work demonstrates that existing LLMs can be effectively extended far beyond their original context windows through careful engineering and clever data synthesis. The method requires only readily available open-source components, making it highly accessible to the research community. The researchers have validated their approach across multiple model sizes (1B, 3B, 8B parameters) and even pushed to 4M tokens in experimental settings.

Just watched a sales leader lose 5 of his top reps after spending months perfecting a "winning" sales methodology that his team HATED. After 18 months of work, the CEO killed his career with six words: "Your team keeps missing their numbers." After analyzing 300+ sales teams and thousands of reps I've identified the exact leadership framework that separates 90%+ quota attainment from the industry average of 60%. The BIG missing piece that most sales leaders miss? Stop running meetings as status updates. And start treating them as PERFORMANCE ACCELERATION ENGINES. Here is the GOLDEN Leadership framework: GROWTH MINDSET: Start every meeting with these 3 strategic elements. → Team member shares industry insight or sales technique (creates learning culture) → Discuss application to current deals (makes learning actionable) → Rotate presenters weekly (builds leadership skills company-wide) This approach increased team knowledge retention by 72% across my client base. OPTIMIZATION SESSION: Have top performers demonstrate and teach these 4 specific skills. → Objection handling techniques (with exact language used) → Discovery questions that uncovered hidden needs → Email templates that generated 80%+ response rates → Closing language that accelerated decisions Use this exact script: "Jeff, you closed that impossible deal with [company]. Walk us through exactly how you handled their [specific objection] so the team can replicate it." LEADERBOARD ACCOUNTABILITY: Create what I call the "Performance Matrix" with columns for. → # of Booked Discovery Calls (activity metric) → New opportunities generated (pipeline metric) → Percentage to monthly target (results metric) → Weekly win or learning (growth metric) DATA & DEVELOPMENT: Each rep inputs and shares three critical elements. → KPIs for the week (leading indicators - 100% controllable) → Sales results (lagging indicators - what they actually sold) → Wins or learnings (development indicators) EXECUTION: Randomly select an AE to role play live. → Use a jar or spinning wheel to pick sales scenarios → Focus on objections, cold calls, or tough situations → Play the difficult prospect yourself → Provide immediate feedback and coaching This gets your team sharper before they jump into their day, and knowing they might be selected drives preparation. NEXT LEVEL MINDSET: End with motivation to conquer the week. → Short visionary speech or gratitude to the team → Positive reinforcement → Ensure they leave with the right mindset This is what they'll remember as they enter their next task or meeting. "REAL RESULTS from this framework: ✅ An IT services client increased sales by 37% in just 30 days ✅ Average rep retention improved from 18 months to 36+ months ✅ Team productivity increased 42% with the same headcount ✅ Top performers stopped taking recruiter calls Hey sales leaders… want a deep dive? Go here: https://lnkd.in/e2iZ7Rmv

𝘓𝘦𝘵'𝘴 𝘩𝘢𝘷𝘦 𝘢 𝘧𝘶𝘯𝘦𝘳𝘢𝘭 𝘧𝘰𝘳 𝘴𝘢𝘭𝘦𝘴 𝘴𝘤𝘳𝘪𝘱𝘵𝘴. ✝️ 𝗥𝗜𝗣 𝗿𝗶𝗴𝗶𝗱 𝘁𝗮𝗹𝗸 𝘁𝗿𝗮𝗰𝗸𝘀: 𝟭𝟵𝟱𝟬𝘀 - 𝟮𝟬𝟮𝟱 ✝️ If your team is still using word-for-word scripts, you're likely seeing: • Prospects who "need to think about it" • Increasingly shorter calls • Objections you can't overcome • Reps who sound robotic and inauthentic Today's buyers can smell a script from a mile away. And they hate it. What's replacing scripts? Conversation frameworks. Working with a B2B services client, we replaced their scripts with flexible frameworks: • Core problem statements (not feature pitches) • Story libraries (not memorized anecdotes) • Question flows (not interrogation lists) • Objection pathways (not canned responses) Results after 60 days: • Average call time increased by 38% • Prospect engagement (measured by questions asked) up 45% • Conversion to next steps improved by 31% The best part? Reps reported feeling "human again" and actually enjoying their calls. Does your team sound like real humans having valuable conversations, or robots reciting memorized lines? If it's the latter, let's talk about building frameworks that actually work. #SalesScripts  #AuthenticSelling  #SalesTraining

I’ve placed 50,000+ candidates using these exact frameworks my students use to land offer letters at top firms. Here are the 5 most common stress-problem interview questions you must prepare, with expert-backed frameworks & concrete examples for each: 1️⃣ “Describe a time you had to make a decision with incomplete information.” Framework: Clarify → Assumptions → Evaluate Options → Choose & Explain Trade-Offs → Validate & Reflect. (Rooted in decision science) Example: As a product analyst, I had 2 days to decide product pricing without regional cost data. I clarified what data I had, stated assumptions about logistics costs, evaluated three pricing models, chose one with buffer margin, and after launch validated real costs. Result: pricing was off by <5%, reducing potential loss by ₹2 lakhs. 2️⃣ “Tell me about when multiple priorities clashed and what did you do first?” Framework: Urgency vs Impact Matrix + Stakeholder Negotiation + Clear Plan. Example: As marketing lead, campaign, content creation, and vendor approvals all due in the same week. I mapped urgency/impact, did vendor first (high impact, low effort), deferred some content with stakeholders, delegated minor tasks. We met major deadlines, revenue targets, without burnout. 3️⃣ “Give an example of when someone challenged your solution. How did you respond?” Framework: Present Solution → Invite Criticism → Adjust with Data & Listening → Finalize. Example: In an analytics project, I proposed using one statistical model. A peer challenged my assumptions about data distribution. I rechecked, collected extra data, and adjusted model inputs. Presentation showed both versions; the final version improved prediction accuracy by 12%. Stakeholders accepted an adjusted one. 4️⃣ “When have you had to think on your feet/sudden change?” Framework: Pause → Clarify scope → Rapid Ideation of alternatives → Choose best → Communicate. Example: During presentation, client asked for metrics by region not prepared. I paused, clarified whether broad region suffice, improvised splits based on last quarter with disclaimers, and focused the rest of the deck on what I had strong data for. The client was impressed by composure; I received follow-up work. 5️⃣ “Describe a time you prevented a problem before it became big.” Framework: Early Diagnosis (monitoring) → Root Cause Analysis (5 Whys / issue tree) → Low-effort Action → Monitor Change. Example: In operations, I noticed error rates slowly rising. Used root cause analysis to find misconfiguration in automation script. Fixed script, added automated alert. Errors dropped by 80%. Saved team 10 hours/week in fixes. If this helped you, repost this post with one of your own answers to any of the above 5 questions using one of these frameworks. Tag me and I’ll pick 5 replies and give feedback on structure & clarity so you can sharpen them before your next interview. #interviewtips #stressinterview #behavioralquestions #careergrowth #dreamjob #interviewcoach

Interview Conversation Role: RTE in #SAFe framework Topic: Conflict Management 👴 Interviewer: "Imagine the Product Manager and System Architect disagree over feature priorities, with the PM focusing on customer needs and the Architect concerned about tech debt. As the RTE, how would you handle this?" 🧑 Candidate: "I’d remind them to focus on the PI objectives and find a middle ground." 👴 Interviewer: "Say this disagreement is slowing decision-making, impacting team alignment, and morale is dipping. What specific actions would you take to mediate?" 🧑 Candidate: "I’d encourage both of them to think about the project’s overall goals." What a skilled Release Train Engineer should say: ------------------------------------------------------ In cases like this, it’s crucial to foster open, constructive discussions without losing sight of both customer value and technical stability. 🌟 I’d start by facilitating a conversation with the PM and Architect to unpack their priorities and establish a shared understanding. 📅 In a similar situation, I scheduled a conflict-resolution workshop with both roles, focusing on ‘value vs. sustainability’ using the Economic Framework. 🏹 We assessed the impact of each priority on the PI objectives, assigning weights based on business and architectural needs. The workshop helped clarify the ROI of tech improvements and immediate features, allowing them to make informed trade-offs. 🛠 To make it concrete, we identified one high-priority feature to advance and a critical refactor for the next PI. By reaching a balanced decision, we addressed urgent customer needs while setting a feasible path for addressing tech debt. 🚩 Impact: This approach helped restore team alignment, fostered trust between the PM and Architect, and improved the ART’s overall efficiency. ✍ As an RTE, my role is to mediate these discussions by grounding decisions in shared values and structured prioritization, ensuring both immediate and long-term value are achieved.