Let's Data Science@letsdatascience

Expore . Learn . Achieve

 New Delhi, India
Joined July 2023
"not approved" contains "approved". A substring check passes the wrong LLM label. For this fixed-label contract, trim outer spaces, lowercase, then compare the whole string. letsdatascience.com/learn/ai… #DataScience
LDS worked example for LLM label evaluation. Expected label: approved. Ignore case and outer spaces; extra words fail. APPROVED with outer spaces passes after normalization. not approved and approved later fail exact comparison, but both contain approved. Code strips outer spaces, lowercases, then compares the whole string. Official LDS logo, Let's Data Science and letsdatascience.com appear below.

ALT LDS worked example for LLM label evaluation. Expected label: approved. Ignore case and outer spaces; extra words fail. APPROVED with outer spaces passes after normalization. not approved and approved later fail exact comparison, but both contain approved. Code strips outer spaces, lowercases, then compares the whole string. Official LDS logo, Let's Data Science and letsdatascience.com appear below.

1
7
A model that loads can still run out of memory. Budget for the KV cache and runtime buffers, then test your intended context and concurrency. letsdatascience.com/blog/llm… #DataScience
LDS decision map for LLM memory planning. Exactly 8 billion four-bit weights occupy 4 decimal GB before overhead. Check loaded weights, then room for KV cache and runtime buffers. If either fails, adjust model or workload and retest. If both pass, measure task quality and latency. Official LDS logo, Let's Data Science and letsdatascience.com appear below.

ALT LDS decision map for LLM memory planning. Exactly 8 billion four-bit weights occupy 4 decimal GB before overhead. Check loaded weights, then room for KV cache and runtime buffers. If either fails, adjust model or workload and retest. If both pass, measure task quality and latency. Official LDS logo, Let's Data Science and letsdatascience.com appear below.
12
SQL challenge: two event tables, one contributor list. Include published posts OR published comments, return each user once, then sort by user_id. Which rows survive? letsdatascience.com/problems… #DataScience
SQL contributor challenge with illustrative input. Posts: user 7 published twice, 9 hidden, 12 published. Comments: 7 published, 9 published, 15 hidden, 18 published. Return user_id for anyone with a published post OR comment, once per user, sorted ascending. The answer is withheld. Official LDS logo, Let's Data Science and letsdatascience.com appear below.

ALT SQL contributor challenge with illustrative input. Posts: user 7 published twice, 9 hidden, 12 published. Comments: 7 published, 9 published, 15 hidden, 18 published. Return user_id for anyone with a published post OR comment, once per user, sorted ascending. The answer is withheld. Official LDS logo, Let's Data Science and letsdatascience.com appear below.
10
Integer index labels are still labels. In this pandas example, df.loc[10:30] keeps A, B and C; df.iloc[0:2] keeps A and B. Check what your slice endpoints mean. letsdatascience.com/learn/pa… #DataScience
Pandas slice example with sorted unique index labels 10, 20, 30, 40, positions 0, 1, 2, 3, and values A, B, C, D. df.loc[10:30] returns A, B, C because its stop label is included. df.iloc[0:2] returns A, B because its stop position is excluded. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Pandas slice example with sorted unique index labels 10, 20, 30, 40, positions 0, 1, 2, 3, and values A, B, C, D. df.loc[10:30] returns A, B, C because its stop label is included. df.iloc[0:2] returns A, B because its stop position is excluded. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
10
A nested append and an outer-list append can affect different objects. Predict a and b in this Python copy challenge, then explain each mutation before running it. letsdatascience.com/quick-re… #DataScience
Python recall challenge with an illustrative Python 3 example: a = [[1], [2]]; b = a.copy(); b[0].append(9); b.append([3]). Predict a and b before running the code, and explain which objects each append changes. Blank answer boxes withhold the solution. The footer shows the official LDS logo, Let's Data Science and letsdatascience.com.

ALT Python recall challenge with an illustrative Python 3 example: a = [[1], [2]]; b = a.copy(); b[0].append(9); b.append([3]). Predict a and b before running the code, and explain which objects each append changes. Blank answer boxes withhold the solution. The footer shows the official LDS logo, Let's Data Science and letsdatascience.com.
14
A 10:00 prediction cannot use a 10:05 feature update. Data Engineer proof: a time-safe join test. ML Engineer proof: re-evaluation with corrected features. letsdatascience.com/learn/pa… #DataScience
Illustrative ML feature timeline: value 2 at 09:55 is eligible for a 10:00 prediction; value 7 at 10:05 is future data. Assume immediate availability. Data Engineer portfolio evidence is an as-of join test excluding the future row. ML Engineer evidence is re-evaluation on corrected features. Responsibilities overlap. Official LDS logo, Let's Data Science and letsdatascience.com appear below.

ALT Illustrative ML feature timeline: value 2 at 09:55 is eligible for a 10:00 prediction; value 7 at 10:05 is future data. Assume immediate availability. Data Engineer portfolio evidence is an as-of join test excluding the future row. ML Engineer evidence is re-evaluation on corrected features. Responsibilities overlap. Official LDS logo, Let's Data Science and letsdatascience.com appear below.
18
User 17 gets prompt A, then B, then A. That breaks a one-variant-per-user LLM experiment. Audit assignment across turns before comparing answer quality. letsdatascience.com/learn/ab… #DataScience
LLM prompt experiment assignment check. For an illustrative test with one prompt variant per user, user 17 receives A, B, A across three turns in the broken trace and A, A, A in the stable trace. Audit user ID, experiment ID and prompt version. This checks assignment only and does not identify a winning prompt. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT LLM prompt experiment assignment check. For an illustrative test with one prompt variant per user, user 17 receives A, B, A across three turns in the broken trace and A, A, A in the stable trace. Audit user ID, experiment ID and prompt version. This checks assignment only and does not identify a winning prompt. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
33
Held-out name: Rex. Primary character: Milo. Story: "Milo helped REX cross the river." A primary-name check misses the leak. Scan the whole row before splitting. letsdatascience.com/learn/fi… #DataScience
Illustrative fine-tuning data check: held-out name Rex, primary name Milo, story 'Milo helped REX cross the river.' The primary-name check misses the leak; the whole-row check finds REX and rejects the row. Scan request, story and primary name using whole-word matching that ignores case, before writing splits. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Illustrative fine-tuning data check: held-out name Rex, primary name Milo, story 'Milo helped REX cross the river.' The primary-name check misses the leak; the whole-row check finds REX and rejects the row. Scan request, story and primary name using whole-word matching that ignores case, before writing splits. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
19
Test your AI agent's approval boundary: approve closing ticket 42, then attempt ticket 99. The mock tool must not run under the old approval. Keep the trace. letsdatascience.com/learn/pa… #DataScience
AI portfolio test, labeled conceptual illustration with a mock ticket tool. Approved action: close_ticket(ticket_id=42). Proposed execution: close_ticket(ticket_id=99). The target changed. Expected trace: argument mismatch, block tool call, request new approval. Assert that close_ticket was not called. Keep the approved action, attempted action and tool-call log. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT AI portfolio test, labeled conceptual illustration with a mock ticket tool. Approved action: close_ticket(ticket_id=42). Proposed execution: close_ticket(ticket_id=99). The target changed. Expected trace: argument mismatch, block tool call, request new approval. Assert that close_ticket was not called. Keep the approved action, attempted action and tool-call log. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
20
LoRA has two parallel paths: frozen base weights and a trainable low-rank update. Add their outputs. Fewer trainable parameters do not imply the same reduction in GPU memory. letsdatascience.com/blog/fin… #DataScience
LoRA architecture for one illustrative linear layer. Input x splits into frozen W0, 4096 by 4096, and trainable A, 16 by 4096, then B, 4096 by 16, scaled by alpha/r. The two outputs add to y. A and B contain 131,072 trainable parameters versus 16,777,216 in the full matrix, 128 times fewer for this layer. These counts do not measure total GPU memory or speed. Official LDS logo, Let's Data Science and letsdatascience.com appear below.

ALT LoRA architecture for one illustrative linear layer. Input x splits into frozen W0, 4096 by 4096, and trainable A, 16 by 4096, then B, 4096 by 16, scaled by alpha/r. The two outputs add to y. A and B contain 131,072 trainable parameters versus 16,777,216 in the full matrix, 128 times fewer for this layer. These counts do not measure total GPU memory or speed. Official LDS logo, Let's Data Science and letsdatascience.com appear below.
24
Orders 101 and 102 are unique. Customer 99 is missing. One test passes; another fails. Keep this small counterexample in your analytics engineering portfolio. letsdatascience.com/learn/pa… #DataScience
Analytics engineering comparison using illustrative data. Orders (order_id, customer_id) are (101,7) and (102,99); customers are 7 and 8. Unique order IDs pass because 101 and 102 differ. The customer relationship fails because ID 99 is missing from customers. A passing uniqueness test does not prove a valid relationship. Portfolio task: keep this failing case with your tests. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Analytics engineering comparison using illustrative data. Orders (order_id, customer_id) are (101,7) and (102,99); customers are 7 and 8. Unique order IDs pass because 101 and 102 differ. The customer relationship fails because ID 99 is missing from customers. A passing uniqueness test does not prove a valid relationship. Portfolio task: keep this failing case with your tests. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
16
The closest neighbor says A. The next two say B. With uniform voting, k=1 predicts A and k=3 predicts B. Check the vote before explaining the classifier. letsdatascience.com/learn/ml… #DataScience
Two-panel comic explaining k-nearest-neighbors classification. The closest example has label A, followed by two B neighbors. With uniform weights and no distance ties, k=1 predicts A; k=3 predicts B by two votes to one. The lesson is to check k and the voting rule before explaining a prediction. This is an illustrative example, not an accuracy comparison. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Two-panel comic explaining k-nearest-neighbors classification. The closest example has label A, followed by two B neighbors. With uniform weights and no distance ties, k=1 predicts A; k=3 predicts B by two votes to one. The lesson is to check k and the voting rule before explaining a prediction. This is an illustrative example, not an accuracy comparison. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
14
Raw dot product ranks A first; cosine ranks B first. Same vectors, different metric. Check normalization before changing your embedding index. letsdatascience.com/blog/tex… #DataScience
Illustrative embedding search example: query (1,0), candidate A (3,4), candidate B (2,0). Raw dot scores are 3 for A and 2 for B, ranking A first. Cosine scores are 0.60 for A and 1.00 for B, ranking B first. Unit normalization gives A=(0.6,0.8) and B=(1,0), so dot product equals cosine. Check the model's recommended metric and normalization. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Illustrative embedding search example: query (1,0), candidate A (3,4), candidate B (2,0). Raw dot scores are 3 for A and 2 for B, ranking A first. Cosine scores are 0.60 for A and 1.00 for B, ranking B first. Unit normalization gives A=(0.6,0.8) and B=(1,0), so dot product equals cosine. Check the model's recommended metric and normalization. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
12
Four token IDs give three next-token targets. Check the first and last input-target pair before training: matching each token to itself teaches copying. letsdatascience.com/learn/bu… #DataScience
Next-token training worked example with illustrative token IDs [17, 42, 9, 31]. Input tokens[:-1] is [17, 42, 9]; target tokens[1:] is [42, 9, 31]. Arrows connect each input position to its next-token target. At input position 2, the visible prefix [17, 42] predicts 9. Positions are 1-based, with no start or end tokens added. Check the first and last pair before training. Official LDS logo, Let's Data Science and letsdatascience.com appear below.

ALT Next-token training worked example with illustrative token IDs [17, 42, 9, 31]. Input tokens[:-1] is [17, 42, 9]; target tokens[1:] is [42, 9, 31]. Arrows connect each input position to its next-token target. At input position 2, the visible prefix [17, 42] predicts 9. Positions are 1-based, with no start or end tokens added. Check the first and last pair before training. Official LDS logo, Let's Data Science and letsdatascience.com appear below.
23
Six shuffled sessions, two users. Calculate each user's two-session rolling average, then shuffle the input again. Does every session keep its result? letsdatascience.com/problems… #DataScience
Python challenge: keep each user's history separate. Illustrative playback_sessions rows (session_id, user_id, watch_seconds): (4,102,300), (1,101,40), (6,102,120), (3,101,60), (2,102,900), (5,101,20). Within each user, sort by session_id and average the current and previous session. Use NaN until that user has two sessions. Return rows sorted by user_id then session_id. What should user_rolling_avg contain? The official LDS logo, Let's Data Science and letsdatascience.com appear below.

ALT Python challenge: keep each user's history separate. Illustrative playback_sessions rows (session_id, user_id, watch_seconds): (4,102,300), (1,101,40), (6,102,120), (3,101,60), (2,102,900), (5,101,20). Within each user, sort by session_id and average the current and previous session. Use NaN until that user has two sessions. Return rows sorted by user_id then session_id. What should user_rolling_avg contain? The official LDS logo, Let's Data Science and letsdatascience.com appear below.
21
Three customers, one paid order. A paid filter in ON keeps all customers; the same filter in WHERE keeps only the paid match. Check which population your report needs. letsdatascience.com/learn/sq… #DataScience
LEFT JOIN worked example with illustrative data. Customers 1, 2 and 3; customer 1 has a paid order, customer 2 a pending order, and customer 3 none. Adding AND o.status = 'paid' to ON returns 1 paid, 2 NULL and 3 NULL. Adding WHERE o.status = 'paid' returns only 1 paid. ON decides matches; WHERE filters joined rows. Official LDS logo, Let's Data Science and letsdatascience.com appear below.

ALT LEFT JOIN worked example with illustrative data. Customers 1, 2 and 3; customer 1 has a paid order, customer 2 a pending order, and customer 3 none. Adding AND o.status = 'paid' to ON returns 1 paid, 2 NULL and 3 NULL. Adding WHERE o.status = 'paid' returns only 1 paid. ON decides matches; WHERE filters joined rows. Official LDS logo, Let's Data Science and letsdatascience.com appear below.
1
20
Same four learners, two quiz scores each. Paired or unpaired t-test? Assume the relevant test assumptions hold, then name your unit of analysis before choosing a function. letsdatascience.com/quick-re… #DataScience
Statistics recall challenge: the same four learners take a quiz before and after training. Illustrative scores out of 100: A 60 to 68, B 80 to 81, C 70 to 76, D 90 to 92. Choose a paired or unpaired t-test and name the unit of analysis. Assume the relevant t-test assumptions hold. The footer shows the official LDS logo, Let's Data Science and letsdatascience.com.

ALT Statistics recall challenge: the same four learners take a quiz before and after training. Illustrative scores out of 100: A 60 to 68, B 80 to 81, C 70 to 76, D 90 to 92. Choose a paired or unpaired t-test and name the unit of analysis. Assume the relevant t-test assumptions hold. The footer shows the official LDS logo, Let's Data Science and letsdatascience.com.
21
One AI reranker, two tests: compare relevance with reranking on/off, then test serving latency under load. Applied science and ML engineering emphasize different evidence. letsdatascience.com/learn/pa… #DataScience
One AI reranker, two tests. Applied Scientist: compare relevance with reranking on versus off, keeping queries, candidates and relevance labels fixed; produce a quality comparison and failure examples. ML Engineer: compare low and high request load with the same model version; report tail latency and errors. This is an illustrative project and responsibilities overlap. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT One AI reranker, two tests. Applied Scientist: compare relevance with reranking on versus off, keeping queries, candidates and relevance labels fixed; produce a quality comparison and failure examples. ML Engineer: compare low and high request load with the same model version; report tail latency and errors. This is an illustrative project and responsibilities overlap. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
16
Two LLM runs, same mean correct-token probability: 0.50. Their losses differ: 1.204 versus 0.693 nats. Take the negative log per token before averaging. Test your scorer with this pair. letsdatascience.com/learn/ll… #DataScience
LLM metric check with two illustrative token positions. Run A assigns correct-token probabilities 0.90 and 0.10; run B assigns 0.50 and 0.50. Both mean probabilities are 0.50, but mean cross-entropy is 1.204 nats for A and 0.693 for B. Take the negative natural log per token before averaging. The table uses the same targets and tokenizer, not measured benchmarks. Official LDS logo, Let's Data Science and letsdatascience.com appear below.

ALT LLM metric check with two illustrative token positions. Run A assigns correct-token probabilities 0.90 and 0.10; run B assigns 0.50 and 0.50. Both mean probabilities are 0.50, but mean cross-entropy is 1.204 nats for A and 0.693 for B. Take the negative natural log per token before averaging. The table uses the same targets and tokenizer, not measured benchmarks. Official LDS logo, Let's Data Science and letsdatascience.com appear below.
1
16
Your LLM call does not get a fresh request budget. With 8 seconds total, 3 already spent and 1 reserved for validation and response, cap the call at 4 seconds. letsdatascience.com/learn/pr… #DataScience
Debugging clinic: an 8-second request budget minus 3 seconds already spent and 1 reserved for validation and response leaves at most 4 seconds for the model call. Cap a configured 6-second timeout at 4 seconds. Recompute remaining time before every attempt. These are illustrative budgets, not measured latency. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Debugging clinic: an 8-second request budget minus 3 seconds already spent and 1 reserved for validation and response leaves at most 4 seconds for the model call. Cap a configured 6-second timeout at 4 seconds. Recompute remaining time before every attempt. These are illustrative budgets, not measured latency. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
1
21
A RAG search that omits D7 does not prove D7 was deleted. Track its chunk IDs, remove them from the index, then verify each ID is absent. Make that your portfolio evidence. letsdatascience.com/learn/pa… #DataScience
Conceptual RAG portfolio example. Source D7 maps to chunks D7:1, D7:2 and D7:3. Retire D7 by deleting every mapped chunk from the index. Capture that the IDs exist before removal and each is absent afterward. Missing from top-k is not proof of deletion. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Conceptual RAG portfolio example. Source D7 maps to chunks D7:1, D7:2 and D7:3. Retire D7 by deleting every mapped chunk from the index. Capture that the IDs exist before removal and each is absent afterward. Missing from top-k is not proof of deletion. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
13
Processing token 4 helps predict token 5. Its query attends to cached K/V plus its own pair; K4 and V4 stay for the next step. letsdatascience.com/blog/gpt… #DataScience
Single-layer KV-cache diagram. Earlier keys K1 to K3 and values V1 to V3 are reused. Token 4 produces Q4, K4 and V4. K4 and V4 join the cache; Q4 attends to all four key-value pairs. The layer produces token 4 representation, and the remaining model predicts token 5. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Single-layer KV-cache diagram. Earlier keys K1 to K3 and values V1 to V3 are reused. Token 4 produces Q4, K4 and V4. K4 and V4 join the cache; Q4 attends to all four key-value pairs. The layer produces token 4 representation, and the remaining model predicts token 5. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
32
Eight staff available, ten required: a hard minimum is infeasible. A soft target can report a shortfall of two. That changes the model, not the business approval. letsdatascience.com/learn/pa… #DataScience
Illustrative staffing comparison. Need ten staff; eight available. Hard requirement: x <= 8 and x >= 10 has no feasible assignment. Soft target: minimize s subject to x <= 8 and x + s >= 10, with nonnegative integer x and s. Best is eight assigned and two unmet. A shortfall penalty changes the model; it does not authorize understaffing. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Illustrative staffing comparison. Need ten staff; eight available. Hard requirement: x <= 8 and x >= 10 has no feasible assignment. Soft target: minimize s subject to x <= 8 and x + s >= 10, with nonnegative integer x and s. Best is eight assigned and two unmet. A shortfall penalty changes the model; it does not authorize understaffing. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
20
A ticket classifier can serve two deadlines: a prediction during a request, or an overnight report. Choose online or batch inference from that need, then show the timing evidence. letsdatascience.com/learn/pa… #DataScience
MLOps portfolio decision map. Does the consumer need a prediction during the request? Yes: online inference, classify a support ticket as it arrives, build an API with a timeout and show latency under load. If results can wait: batch inference, classify yesterday's tickets overnight, build a scheduled scoring job and show completion time and failed rows. Illustrative portfolio tasks. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT MLOps portfolio decision map. Does the consumer need a prediction during the request? Yes: online inference, classify a support ticket as it arrives, build an API with a timeout and show latency under load. If results can wait: batch inference, classify yesterday's tickets overnight, build a scheduled scoring job and show completion time and failed rows. Illustrative portfolio tasks. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
15
escalate: false. Reason: security. Both fields can pass their individual checks and still conflict. Validate the pair before your LLM workflow routes the ticket. letsdatascience.com/learn/co… #DataScience
Illustrative two-panel LLM workflow comic. An output excerpt pairs escalate: false with escalation_reason: security. Individual field checks pass, but the pair conflicts. An engineer stops routing and requires a reason exactly when escalation is true. Run the application check before routing. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Illustrative two-panel LLM workflow comic. An output excerpt pairs escalate: false with escalation_reason: security. Individual field checks pass, but the pair conflicts. An engineer stops routing and requires a reason exactly when escalation is true. Run the application check before routing. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

1
31
Same ten RAG candidates, one relevant result: moving it from rank 6 to 2 changes Recall@5 from 0 to 1. Recall@10 stays at 1. Keep the cutoff explicit when comparing rerankers. letsdatascience.com/learn/ra… #DataScience
Illustrative RAG reranking example with one query and one relevant result among ten candidates. Before: relevant result at rank 6. After: rank 2. The top-5 cutoff stays between rows 5 and 6. Recall@5 changes from 0 to 1; Recall@10 stays 1; reciprocal rank changes from 1/6 to 1/2. A reranker reorders the shortlist and cannot add a missing result. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Illustrative RAG reranking example with one query and one relevant result among ten candidates. Before: relevant result at rank 6. After: rank 2. The top-5 cutoff stays between rows 5 and 6. Recall@5 changes from 0 to 1; Recall@10 stays 1; reciprocal rank changes from 1/6 to 1/2. A reranker reorders the shortlist and cannot add a missing result. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
13
Call t2 finishes before t1. Your agent must still attach each result to its original request. Match by call ID; test with reversed completion order. letsdatascience.com/learn/bu… #DataScience
Illustrative AI agent trace. Two get_order_status requests run: t1 for order A and t2 for order B. Time flows downward. Amber t2 returns processing first; teal t1 returns shipped later. Each result retains its call ID. Test by reversing completion order and checking both mappings. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Illustrative AI agent trace. Two get_order_status requests run: t1 for order A and t2 for order B. Time flows downward. Amber t2 returns processing first; teal t1 returns shipped later. Each result retains its call ID. Test by reversing completion order and checking both mappings. The official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
19
Top-p = 0.80: if the first two tokens total 0.75, the third stays. The threshold governs token probability, not factual correctness. letsdatascience.com/blog/llm… #DataScience
LLM sampling decision map: Top-K selects a token count; Top-P selects a ranked prefix reaching cumulative probability p; Min-P uses a fraction of the best token probability. Illustrative probabilities A 0.55, B 0.20, C 0.15 and D 0.10: top-p 0.80 keeps A, B and C, totaling 0.90, because A+B is only 0.75. Drop D, renormalize, then sample. Token probability is not factual confidence. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT LLM sampling decision map: Top-K selects a token count; Top-P selects a ranked prefix reaching cumulative probability p; Min-P uses a fraction of the best token probability. Illustrative probabilities A 0.55, B 0.20, C 0.15 and D 0.10: top-p 0.80 keeps A, B and C, totaling 0.90, because A+B is only 0.75. Drop D, renormalize, then sample. Token probability is not factual confidence. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
15
Two delivered shipments share the latest date. Which one survives the tie-break? Use the illustrative rows, then write the full SQL query. letsdatascience.com/problems… #DataScience
SQL challenge with illustrative shipments for order_id 42. Shipment 18: delivered, 2026-09-09; 21: delivered, 2026-09-10; 19: delivered, 2026-09-10; 24: in_transit, NULL. Return one row per purchase. Only delivered rows with a delivery date qualify. Choose the latest date, breaking ties by the smallest shipment_id. Which shipment survives? Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT SQL challenge with illustrative shipments for order_id 42. Shipment 18: delivered, 2026-09-09; 21: delivered, 2026-09-10; 19: delivered, 2026-09-10; 24: in_transit, NULL. Return one row per purchase. Only delivered rows with a delivery date qualify. Choose the latest date, breaking ties by the smallest shipment_id. Which shipment survives? Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
15
Two team averages: 2 and 10 minutes. Their ticket counts: 10 and 90. The combined mean is 9.2 minutes. Averaging the means gives each team equal weight, even though their volumes differ. letsdatascience.com/learn/st… #DataScience
Worked example: combine means with counts. Team A handles 10 tickets averaging 2 minutes, totaling 20 minutes. Team B handles 90 averaging 10 minutes, totaling 900. The combined mean is 920 divided by 100, or 9.2 minutes per ticket. Averaging 2 and 10 gives 6 and weights the teams equally. Illustrative data assumes the same metric and period with no overlapping tickets. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Worked example: combine means with counts. Team A handles 10 tickets averaging 2 minutes, totaling 20 minutes. Team B handles 90 averaging 10 minutes, totaling 900. The combined mean is 920 divided by 100, or 9.2 minutes per ticket. Averaging 2 and 10 gives 6 and weights the teams equally. Illustrative data assumes the same metric and period with no overlapping tickets. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
16
Sample size rises from 100 to 400. Both samples have SD = 20. For independent observations, what happens to the standard error of the mean? Calculate it before checking the distinction. letsdatascience.com/quick-re… #DataScience
Statistics recall challenge. Assume independent observations from the same population. Sample A has n=100 and sample SD=20. Sample B has n=400 and sample SD=20. SE of the mean equals sample SD divided by the square root of n. Calculate both standard errors and explain whether a more precise mean implies less spread in individual values. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.

ALT Statistics recall challenge. Assume independent observations from the same population. Sample A has n=100 and sample SD=20. Sample B has n=400 and sample SD=20. SE of the mean equals sample SD divided by the square root of n. Calculate both standard errors and explain whether a more precise mean implies less spread in individual values. Official LDS logo, Let's Data Science and letsdatascience.com appear in the footer.
9
Introducing Let's Data Science — SQL & Python practice on 15 real industry domains. 1,500+ problems. Browser-based. Instant grading. 1,000 founding member spots at $3.99/mo. letsdatascience.com/pricing
2
28
Load more