| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
fix(data-management): update canonical hugging face dataset paths and configs (#180) * fix: update dataset path for Rotten Tomatoes in load_and_inspect and stream_dataset functions * fix: improve formatting of dataset split print statements * fix: update Hugging Face IDs for dataset recommendations in prompt-data-helper * fix: update Hugging Face IDs and configurations in dataset recommendations | 2 个月前 | |
feat(site): full curriculum figure coverage (503 lessons) + lang-picker CSS cachebust (#380) * feat(site): 39 more interactive figures — wave 3 across thin phases Extends the figure system into the phases that were still sparse: LLM engineering, multimodal, agents depth, alignment, plus vision/speech/genai remainders. Five new module files (1,959 LOC) on the shared LF toolkit: - figures-llmeng.js (P11/P13, 8): few-shot curve, chain-of-thought, constrained decoding, prompt-cache hit, semantic cache, function-call args, LLM-judge rubric, lost-in-the-middle - figures-multimodal.js (P12, 7): contrastive matrix, cross-attention fusion, modality projection, CFG guidance scale, VQ codebook, video patches, CTC align - figures-agents2.js (P14/P16, 8): ReWOO plan, tree-of-thoughts, self-refine, memory blocks, Voyager skills, LangGraph state, orchestration patterns, debate - figures-alignment2.js (P9/P18, 8): PPO clip, reward model, constitutional AI, actor-critic, interpretability probe, SAE features, jailbreak defense, scalable oversight - figures-foundations2.js (P4/P6/P8, 8): augmentation, transfer learning, BN train/eval, CTC collapse, MFCC pipeline, autoencoder bottleneck, normalizing flow, score matching Embedded in 39 figure-free lessons. Validated headless: all 173 registered figures (16 core + 157 module) mount with zero console errors; PPO clip, CLIP contrastive matrix, and tree-of-thoughts verified in light and dark. * feat(site): 81 animated SVG figures across capstone, agents, CV, NLP, infra Wave 4a. Shift from slider widgets to unique, concept-specific SMIL-animated SVG illustrations (no JS loops, no real compute, light DOM). 10 new module files, each figure a distinct visual matched to its lesson: - figures-capstone-a/b (P19, 16): tokenizer merges, sliding window, training loop, DPO, RAG flow, eval grid sweep, sandbox runner, safety checkpoints - figures-agents3 (P14, 8): HTN tree, workflow chain, actor mailbox, debate convergence, computer-use cursor, voice pipeline, injection hijack, cascade - figures-nlp3 (P5, 9): POS tags, dependency arcs, QA span, summarize collapse, topic drift, coref links, NLI router, relation triples, constrained decode - figures-cv2 (P4, 8): detection NMS, segmentation flood, GAN, diffusion denoise, NeRF rays, CLIP matrix, metric embedding, depth sweep - figures-llms3 (P7/P10, 8): MoE routing, encoder-decoder, RNN vs parallel, speculative draft-verify, multi-token predict, self-critique, loss masking, activation recompute - figures-autonomous2 (P15, 8): AlphaEvolve loop, Darwin-Godel archive, bounded gates, circuit breaker, checkpoint replay, cost governor, injection boundary - figures-swarms2 (P16, 8): consensus wave, auction, stigmergy, hierarchy token, message bus, roles, blackboard, speaker election - figures-infra2 (P17, 8): cache-aware router, cold start, model cascade, prefill/decode split, batch lanes, semantic cache, edge bandwidth, load waves - figures-systems3 (P11/12/13/8/6, 8): masked diffusion, any-to-any stream, video diffusion, inpaint, agentic RAG, MCP NxM, A2A lifecycle, RVQ codec Embedded in 80 figure-free lessons. Validated headless: all 254 registered figures mount with zero console errors, 1062 SMIL animation nodes present, samples (NMS, stigmergy, RAG, detection) verified rendering. * feat(site): 80 more animated SVG figures across capstone, CV, speech, tools Wave 4b. Ten more SMIL-animated module files, each figure a unique concept-specific illustration (no JS loops, no real compute, light DOM): - figures-capstone-c/d (P19, 16): embedding lookup, transformer block, GPT assembly, weight remap, grad accumulation, atomic checkpoint, HyDE, BLEU, reliability diagram, ZeRO shard, pipeline bubble, constitution loop - figures-cv3 (P4, 8): RoIAlign, latent compression, CTC, pose heatmap, gaussian splat, rectified flow, open-vocab, track association - figures-speech2 (P6, 8): ASR attention, EER crossover, TTS stack, codec tokens, VAD cascade, full-duplex, WER alignment, voice factorize - figures-multimodal2 (P12, 8): patch-n-pack, LLaVA projector, M-RoPE axes, video token budget, action tokens, doc layout, MaxSim, agent loop - figures-tools2 (P13, 8): tool loop, parallel fanout, schema routing, client merge, transport handshake, task lifecycle, tool poisoning, router failover - figures-agents4 (P14, 8): memory fusion, crew-vs-flow, handoff, subagent isolation, SWE-bench gate, agent-human gap, span tree, eval layers - figures-swarms3 (P16, 8): contract-net, work-stealing, handoff routing, agent-card discovery, debate topology, theory-of-mind, CTDE, checkpoint - figures-genai3 (P8/P5, 8): VAR next-scale, FID, PatchGAN, StyleGAN mapping, hybrid retrieval, Matryoshka, entity linking, needle-in-haystack - figures-misc2 (P15/P17/P11, 8): propose-then-commit, priority tiers, research loop, speculative tree, gateway fallback, sequential test, schema funnel Embedded in 80 figure-free lessons. Validated headless: all 334 registered figures mount with zero console errors, 2061 SMIL animation nodes; samples (contract-net auction, TTS stack) verified rendering. * fix(site): bump asset versions so the language picker CSS refreshes The picker markup and CSS shipped, but the style.css link kept the old ?v=20260525a query, so returning visitors' browsers served cached CSS without the .lang-panel rules and the picker rendered unstyled and always-open. Bump every asset version (and version the new langs.js / lang-picker.js) to force a fresh fetch. * feat(site): full curriculum figure coverage — 169 animated figures, all 503 lessons Wave 5 completes the interactive figure system: every lesson in every phase now carries a concept-specific animated SVG. Sixteen new module files (7,688 LOC), each figure a unique SMIL illustration with motion craft applied throughout (spline ease-out entries from opacity 0 at 95 percent scale, staggered cascades, exits faster than entries, calm 2.5-6s loops, no JS animation loops, no real compute): - figures-capstone-e/f/g/h/i (P19, 49 figures) - figures-alignment3/4 (P18, 23) - figures-workbench (P14, 15): the agent workbench mini-track animated - figures-tools3 (P13, 11) - figures-setup (P00, 12): commit DAG, GPU dispatch, secret injection, venv isolation, docker layers, LSP round trip, flame graph, more - figures-foundations3 (P01/02/09, 11) - figures-visaudio4 (P04/06/08, 9) - figures-nlp5 (P05/07, 8) - figures-llmstack5 (P10/11/12, 11) - figures-autoswarm5 (P15/16, 12) - figures-infra4 (P17, 8) Coverage: 0 figure-free lessons remain; all 503 lesson docs carry a figure. Validated headless: 506 registered figures mount with zero console errors, 4,505 SMIL animation nodes. | 1 个月前 | |
feat: book edition pipeline, fundamentals-first headlines, animated figures, verified bug fixes (#348) Book pipeline: six-volume EPUB/PDF compilation built by CI from lesson sources (book/, scripts/build_book.py, themed title pages with edition stamps, site-matching print theme), attached to every GitHub release. Homepage Books section and README section link the latest release. Fundamentals-first headline policy across the course: 16 lesson titles and 30+ taglines/section headings now lead with the concept (agent state machines, actor model, role-based teams, memory paging, serving engine internals, permission modes); framework and product names are demoted to attributed in-body examples. README, ROADMAP, quizzes, and prerequisite references synced. New agent-memory taxonomy section maps memory types to representative implementations. Vendor-neutral model policy: runnable defaults read the LLM_MODEL env var with undated aliases; dated snapshot ids removed; multi-provider phrasing in the setup lesson. Lessons deepened with original material: prediction-game origins of perplexity (05/16), scripted-era chatbot lineage 1950-2001 (05/17), causal-triangle derivation from prefix averaging plus GPT-5 date fix (07/07). Three new animated site figures back them (figures-history.js). llms.txt now carries per-lesson raw markdown links so agents can fetch full lesson text directly. Bug fixes verified with executed repros: capstone solved flag keyed to test results, 405B cost estimator overflow, f-string crash on Python <3.12, no-torch demo path, negative stable BCE, all-zero stationary distribution, inverted Cohens d, per-lesson quiz panel, lesson-fetch retry with honest errors, decision-trees doc completed, editor shortcuts, rustc run command, Docker python3.12 build with doc sync, git lesson fork flow, FIPA receiver field, fnm under Rosetta, 15 curl-verified link fixes, remaining imdb dataset id spot. | 1 个月前 | |
fix(quiz): correct answer is always in the same position (slot B) (#381) Every "Test Your Understanding" quiz placed the correct answer in option B. Across the 2026 questions in 338 quiz files the correct answer sat at index 1 in 61.5% of cases (uniform would be ~25%), and 107 files had every answer at B, making the quizzes guessable without reading them. scripts/debias_quizzes.py rewrites each question's option order with a deterministic, content-seeded permutation and updates the correct index to follow the moved answer. It is idempotent: options are canonicalised to a sorted base before permuting, so re-running produces byte-identical output. Questions whose options reference each other by position ("all of the above", "both A and B") are left untouched. The correct-answer value, the option set, and every explanation are preserved exactly; only order and the index change. Result: A 23.8% / B 26.3% / C 23.5% / D 26.4%. The script doubles as a CI guard: `--check` exits non-zero if any quiz is not de-biased, wired into the curriculum workflow so new lessons cannot regress. Fixes #368 | 1 个月前 |