Hebrish: English-Hebrew code-switching
Collection
Every Hebrish model, dataset and Space, including those still on danielrosehill. hebrish.app • 35 items • Updated
id string | text string | target_terms list | categories list | homophone_set null | split string | derivation string | sentence_author string | source_dataset null | source_licence null | source_url null | source_authors list | source_date null | original_text null | review_note string | term_origins list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
p_00919fd6cd | Turn on flash attention so the long recordings fit in memory. | [
"flash attention"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_00f96c9a66 | Open the model picker and switch to Sonnet for the rest of this. | [
"model picker"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_017d980a28 | Idea for the newsletter, a sober piece on what AGI timelines actually claim. | [
"AGI"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_01ad4653eb | Switch the output style to something shorter, I don't need the explanations. | [
"output style"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_0230374116 | I'm skipping the superintelligence debate and focusing on shipping useful tools. | [
"superintelligence"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0238c94375 | Add the GitHub CLI to the allow list for this project only. | [
"allow list"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_02c9d374c2 | Spin up a background agent to sweep the repos for the old API key name. | [
"background agent"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_0301c97351 | Estimate whether this repo fits in the context length of the model I'm using. | [
"context length"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_031927810c | Everyone has a different definition of AGI, which makes the debate pointless. | [
"AGI"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_03d4f59e0e | The model's tool use looked sensible, it read the file before editing it. | [
"tool use"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_05093eb0fe | Tell Cline to stop after the plan and wait for me. | [
"Cline"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_054dd61a19 | We are exploring where a large language model can reduce the support workload. | [
"large language model"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_05e69c1c6e | Before we ship the new prompt, can someone run the evals and post the numbers? | [
"evals"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_06a2eb1328 | Run the same task on GPT-5 Codex and see if it finds the bug faster. | [
"GPT-5 Codex"
] | [
"tech",
"ai_interactions",
"numeric"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_075b7c8aaa | I turned on JSON mode and the model still left out a required key. | [
"JSON mode"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0790d1f4ba | You're the orchestrator on this one, so break it into pieces and hand them out. | [
"orchestrator"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_07aac51321 | We cloned a voice for the text-to-speech demo, and the client was very impressed. | [
"text-to-speech"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_07b79b6733 | Turn on branch protection for master so nobody pushes to it directly. | [
"branch protection"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_0885c0b967 | Our red teaming turned up a prompt that made the bot leak its instructions. | [
"red teaming"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_08b0a5919c | The Stop hook keeps sending you back to work, is it stuck in a loop? | [
"Stop hook"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_09b1a08084 | Strip the outdated tool notes out of the system prompt before the next test. | [
"system prompt"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | fixed: Comma splice and overlaps with existing bank sentences about keeping system prompt short / eating the window. | [
"claude"
] |
p_09b1ee74d2 | Add a timeout inside the agent loop so a hung command cannot block the worker. | [
"agent loop"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | fixed: Textbook definition copying pack meaning. | [
"daniel"
] |
p_0a07256d04 | We would prefer an open source model so the client can host it themselves. | [
"open source model"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0a4feecf1e | Make a custom output style that skips the recap at the end. | [
"output style"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_0a96b9f0f7 | Extract the people and companies from this article and build a small knowledge graph. | [
"knowledge graph"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0b59559140 | Reminder to look into reinforcement learning for the tool use problem. | [
"reinforcement learning"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0c29d7cce6 | Generate the changelog from the commit messages since the last tag. | [
"changelog"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_0c922cdc64 | An open source model is fine for the prototype, and we can revisit later. | [
"open source model"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0d2a254353 | The training crash happens during backpropagation, which points to a memory problem. | [
"backpropagation"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0dbe770870 | Results from the batch API take hours but cost less. | [
"batch API"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0de3214729 | Run it in headless mode from the cron job and log the output to a file. | [
"headless mode"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_0e0b98829d | Re-run the failed workflow run and watch it this time. | [
"workflow run"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_0e3c5cb8eb | Write a short script using the Transformers library to load the model and run one sample. | [
"Transformers library"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0e885c8ebd | Remind me to explain backpropagation in the workshop without any heavy math. | [
"backpropagation"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_0ee7e75789 | Did Roo Code leave any uncommitted changes in this repo? | [
"Roo Code"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_0efc739001 | I checked the bill and prompt caching saved more than the model switch did. | [
"prompt caching"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_0fe3c67993 | Run git worktree list and tell me which ones are stale. | [
"git worktree"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_1003a56f3f | The PreToolUse hook is exiting with an error, so read its log and fix it. | [
"PreToolUse hook"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_102393e48c | CI is red on master, find out which commit broke it. | [
"CI"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_1080edfe2e | Ask Sonnet to find the Shakespeare sonnet she read at the wedding, I want the text for the card. | [
"Sonnet"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_11210dfefb | Can you explain the attention mechanism in two sentences for the product team? | [
"attention mechanism"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_119ad5eb3b | Look at what each attention head does on the long prompts. | [
"attention head"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_11db66a2c8 | The recording app uploads each take to staging once the phone has a connection. | [
"recording app"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_1227eaec3d | Check whether the chat template adds the special tokens before the user turn. | [
"special tokens"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_126efb5ea8 | Create a release tag for this version and push it. | [
"release tag"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_1331b6c8dd | Can you add a section to AGENTS.md about the deploy steps? | [
"AGENTS.md"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_14a7f8389b | Add a pre-commit hook that blocks any file with a token in it. | [
"pre-commit hook"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_14dbc2d804 | Register the new MCP server on the gateway, not directly in Claude Code. | [
"MCP server"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_14e462f0c1 | Which MCP tool did you use to read that file? | [
"MCP tool"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_1530473aca | I'm skeptical that AGI is close, but the coding progress is hard to ignore. | [
"AGI"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_158ce51c72 | Switch to beam search for the final transcripts, since it keeps several candidates alive. | [
"beam search"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_15bfbed21f | Our guardrails are too loose, it approved a reply that quoted private phone numbers. | [
"guardrails"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_15ce822b52 | Mark the draft PR ready once the build passes. | [
"draft PR"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_168378ef9c | You're in detached HEAD, so make a branch before you commit anything. | [
"detached HEAD"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_16d0875104 | I put a quick Gradio demo together so the client can try the model. | [
"Gradio"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_170baf7f55 | Long noisy sessions cause context rot, so I keep each task in its own chat. | [
"context rot"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_17195141a3 | Have Haiku write me a short haiku about merge conflicts for the team chat. | [
"Haiku"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_17421cf09e | Why did the preview deployment fail when master built fine? | [
"preview deployment"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_17b2bbc3bd | Did the Explore agent check the tests folder as well? | [
"Explore agent"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_18042c58fe | Retry the tool call once, and if it fails again, change route. | [
"tool call"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_18383478b9 | The branch has no upstream yet, so the push failed. | [
"upstream"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_1852d2329b | Cap the agent loop at ten steps so a bad run cannot spin forever. | [
"agent loop"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_187530837b | Classify these tickets zero-shot, and tell me which ones you're unsure about. | [
"zero-shot"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_18eddadafb | Ask for chain of thought on the migration, I want to see each step. | [
"chain of thought"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_18ff72f628 | A knowledge graph helped the assistant answer questions that span several documents. | [
"knowledge graph"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_18ffefe80d | Keep AGENTS.md short, because agents skim it at the start of every session. | [
"AGENTS.md"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_1905744091 | The talk on edge AI covered quantization tricks that I want to try. | [
"edge AI"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_19ed2d09f8 | Which foundation model are we fine-tuning for the voice project? | [
"foundation model"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_1a15cc3a40 | Please amend the last commit, you left the debug line in. | [
"amend"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_1ab5c5bad2 | Which version of the SDK is pinned in this project? | [
"SDK"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_1d093ec79e | Ask it to quote the line, because a made-up quote is a hallucination it cannot back up. | [
"hallucination"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_1d1769035a | Who did the force push that wiped yesterday's commits? | [
"force push"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_1e87cfd7d3 | Running backpropagation through the frozen layers wastes memory, so detach them. | [
"backpropagation"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_1fdbdb6a9a | Send the file path instead of dumping raw output into the context window. | [
"context window"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | fixed: Overlaps with existing bank sentence ("Don't paste the whole log in, it'll eat the context window."). | [
"claude"
] |
p_200cfbfed8 | I want to spend an hour on prompt engineering before we touch the training code. | [
"prompt engineering"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_203da7ffbb | Compact the conversation summary now so context rot does not ruin the rest of the refactor. | [
"context rot"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | fixed: Duplicates "Start a fresh session before..." opener from row 54 and "thread" from row 105. | [
"daniel"
] |
p_206f690fb0 | Why didn't ToolSearch return the GitHub tools? | [
"ToolSearch"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_2095ba529d | Codex and OpenCode both read AGENTS.md, so one file covers the whole team. | [
"AGENTS.md"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | fixed: the pack's meaning listed Cursor as a reader of AGENTS.md, which is unverified; the sentence now names the two harnesses that verifiably read it | [
"claude"
] |
p_20b1713716 | I reported a crash in the recording app after it lost audio during a long take. | [
"recording app"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_20c98474b6 | Constitutional AI uses written principles so the model can critique its own answers. | [
"Constitutional AI"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_20d2c69d15 | The changelog is missing the last three versions. | [
"changelog"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_2137175aa6 | Codex left a half-finished branch last night, so check what it changed before we merge anything. | [
"Codex"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_22a2820b4f | Don't give that subagent write access, it only needs to search the files. | [
"subagent"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_23f5d7f6ed | One attention head seems to track the previous token in every layer. | [
"attention head"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_256c798a77 | Is there a skill for this already, or are you improvising the workflow? | [
"skill"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_25c8c537e6 | Does Codex read the same AGENTS.md file, or does it need its own instructions? | [
"Codex"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_266f93d179 | The paper replaces self-attention with a cheaper operation for long inputs. | [
"self-attention"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_26e713fdcd | Split the project rules by folder so the frontend notes stay separate from backend code. | [
"project rules"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | fixed: Repeats "Keep <file> short so the agent reads it" formula from rows 5 and 22. | [
"daniel"
] |
p_273ca4bdb3 | Why is the env file not in the gitignore? | [
"gitignore"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_275e3b9421 | Before compaction kicks in, save the list of open tasks somewhere you can read back. | [
"compaction"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_27c0e4e099 | Get me out of detached HEAD without losing those two commits. | [
"detached HEAD"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_27e53e69e9 | Numerical overflow in the softmax is why we subtract the maximum first. | [
"softmax"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_2835ae3f17 | Load the MCP tool first with ToolSearch, then call it. | [
"MCP tool"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
p_292b0d6ced | Validation loss is rising while training loss falls, which is classic overfitting. | [
"overfitting"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_29d18fbe4c | Write this up so the next Claude Code session can pick it up without asking me anything. | [
"Claude Code"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_2a2aec0f82 | Explain rotary positional encoding in plain terms and show a tiny code example. | [
"positional encoding"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_2a5cefa7e7 | Running batch inference over the whole archive will be far quicker than looping through files. | [
"batch inference"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_2b44311fbb | Strip untrusted markup from incoming emails so a message cannot sneak in prompt injection. | [
"prompt injection"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | fixed: Inverted logic in original ("not as possible prompt injection" implies trusting the email). | [
"daniel"
] |
p_2bc1831ebe | The attention mechanism is the reason these models handle long-range dependencies. | [
"attention mechanism"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"daniel"
] |
p_2be809c21d | Add the new server to the MCP gateway and restart it. | [
"MCP gateway"
] | [
"tech",
"ai_interactions"
] | null | train | generated | claude (Claude Code subagent) | null | null | null | [] | null | null | null | [
"claude"
] |
End of preview. Expand in Data Studio
AI Interactions Sentence Bank
English sentences in the register of instructions dictated to AI coding agents (Claude Code, Codex, Cursor and others) during a working day: subagents, background agents, worktrees, context and compaction, plan mode, hooks, MCP servers, skills and plugins, plus the git and GitHub workflow around them (branches, pull requests, rebases, merge conflicts, GitHub Actions).
Website: hebrish.app: get the Hebrish keyboard for Android and the desktop apps, see the results, and find every model, dataset and demo in one place.
These are sentences to read aloud for fine-tuning a speech-recognition model, not recordings. They feed a targeted fine-tune of NVIDIA Parakeet TDT 0.6B v2 (current model). A capture app shows these sentences one at a time to be read aloud; every take lands in Read-Aloud-Voice-Sentences and is copied into the matching audio subset. The recordings of these sentences are in AI-Interactions-Sentences.
This dataset gathers sentences from several places into one file for my own training, and it is reproduced here for convenience. Every row says where it came from, who wrote it and how it was changed. Credit for upstream sentences belongs to their authors (see "Attribution").
Size
607 train, 70 test (677 total). The test split is term-disjoint:
no target term in test appears in train, so an evaluation on it measures generalisation,
not memorisation.
Derivation: generated 677.
Sentence authors: claude (Claude Code subagent) 677.
Columns
| Column | Meaning |
|---|---|
id |
sentence id: a hash of the text (same id as the capture app and the audio datasets' prompt_id) |
text |
the sentence, exactly as read. Punctuated and cased; numbers in digits for versions, sizes, percentages and money |
target_terms |
the words the sentence exists to teach |
term_origins |
for each target term, who proposed it: daniel (from my own repos and notes, or I asked for it), claude (an AI agent proposed it), or the public list it came from (openrouter, github-topics, cncf, pypi, stackoverflow) |
categories |
tech, linux, smart_home, homophone, numeric, … |
homophone_set |
homophone rows only: the set the context disambiguates |
split |
train or test |
derivation |
mined-verbatim (taken unchanged from a public dataset), mined-edited (taken, then lightly fixed for casing, digits or readability), generated (written by an LLM for this project, then reviewed by a second model), borrowed (taken whole from an earlier sentence set) |
sentence_author |
the upstream dataset, or the model that wrote the sentence |
source_dataset, source_licence, source_url, source_authors, source_date |
upstream provenance for mined and borrowed rows. source_url links the paper or post; source_authors lists arXiv authors (Reddit usernames are deliberately not copied) |
original_text |
the upstream text before editing (mined-edited rows) |
review_note |
the reviewer's note when a sentence was fixed |
Attribution and licence
Licensing is per row (source_licence). Rows without one were written for this project and
are CC-BY-4.0.
| Upstream | Licence |
|---|---|
| (none: every row was written for this project) | CC-BY-4.0 |
Mined sentences were selected by scripts/mine_tech_sentences.py in the project repository:
they are 8–30 words long, contain at least one named tech term, and come from newer posts first.
An AI reviewer then kept or dropped each one and fixed casing and numbers.
Updating
Upstream sources can be re-mined, and new sentences are added as the project grows. Ids are content hashes, so an edited sentence gets a new id.
- Downloads last month
- 116