Datasets:

keppy
/

evalroute-tasks

id stringlengths 14 14	text stringlengths 25 489	label stringclasses 4 values	source stringclasses 2 values	contributor stringclasses 1 value	week stringclasses 2 values
keppy:ddbe8715	Hermes operations bring-up: create a kanban board, clone a worker profile, run the web dashboard bound to a Tailscale IP, create a script-only cron job, verify each step with CLI output. Configuration of shipped Hermes features, no new code.	routine-coding	ledger-pinned	keppy	2026-W40
keppy:66fe5b04	Add three argparse flags to an existing Python CLI (rate --model/--effort forwarding, route --replace-route-id, route --json reusing an existing JSON envelope), fix an env-var fallback in a path helper, add pytest tests following the repo's existing CLI test pattern. Small diff, tests check the result.	routine-coding	ledger-pinned	keppy	2026-W40
keppy:492154b2	One-line Python fix: only append an optional CLI flag to a string when its value is set; add two pytest cases for the uncovered path.	routine-coding	ledger-pinned	keppy	2026-W40
keppy:f9f7fec9	Add a new CLI subcommand module to a Python plugin: compose two existing functions (route lookup, subprocess spawn of a sibling CLI with flags from the lookup), capture stdout/stderr to files, timeout handling, print three summary lines; pytest with a stub executable via env var.	routine-coding	ledger-pinned	keppy	2026-W40
keppy:ebba9855	Correctness review of a 450-line Python diff (CLI plugin: subprocess spawn with timeout on Windows, importlib loading, argparse wiring, ledger writes). Trace code paths against the source; produce structured findings JSON.	dl-ml-research-engineering	ledger-pinned	keppy	2026-W40
keppy:549ed358	Hygiene review of a Python diff against two written briefs: scope creep, test coverage per claim, stray files, stale docs; structured findings JSON.	routine-coding	ledger-pinned	keppy	2026-W40
keppy:963ff684	Apply four small reviewed fixes to a Python CLI plugin (fail-fast on None id, epilog line, one isolation test, Windows process-tree kill on timeout with a grandchild test), then release prep: CHANGELOG entry, version bump in two files, README paragraph. Tests check the result.	routine-coding	ledger-pinned	keppy	2026-W40
keppy:bc30efea	compare two training runs	dl-ml-research-engineering	ledger-pinned	keppy	2026-W40
keppy:cf97ef63	Python CLI library: add a templated --runner flag with shlex-safe placeholder substitution and named presets, write a JSON sidecar file after each run and have the report read it, add --json output to four subcommands with snapshot tests for the human output, add an env var to a path resolver, write AGENTS.md and llms....	routine-coding	ledger-pinned	keppy	2026-W40
keppy:d3df4fc6	Code review for correctness of a Python library refactor and CLI feature set: trace contract guard, subprocess template substitution, JSON sidecar writes on all exit paths, sqlite read-only ingestion, env precedence; report findings as JSON	dl-ml-research-engineering	ledger-pinned	keppy	2026-W40
keppy:ada2e9c9	Code review for hygiene: diff vs brief scope, tests match names, no stray files, README and CHANGELOG claims true, versions untouched; report findings as JSON	routine-coding	ledger-pinned	keppy	2026-W40
keppy:18a6bb58	Role-play a coding agent with no prior context reading a repo's README and AGENTS.md cold, following the documented five-command loop in a fresh venv, and reporting every friction point and doc error as JSON	prose	ledger-pinned	keppy	2026-W40
keppy:e19b1c23	Apply a code-review fix-set to a Python CLI library: thread a surface setting through string helpers so command hints differ by invocation surface with snapshot tests preserved, strip quotes per token after template substitution, fold a printed line into a JSON envelope, make a subcommand exit non-zero on no-op, fix a ...	routine-coding	ledger-pinned	keppy	2026-W40
keppy:fcdc1e9c	Python library feature in three commits: whitelist-redaction of JSONL ledger rows with HMAC task hash, salt and cursor files, a layered opt-in config gate, dry-run CLI output; an upload verb using a monkeypatched huggingface_hub client with byte-identical content test; aggregation of pooled rows with a merge-invariant ...	routine-coding	ledger-pinned	keppy	2026-W40
keppy:5a8ef8ae	fix the failing test in tests/test_x.py	routine-coding	ledger-pinned	keppy	2026-W40
keppy:1b1b6a29	Python library polish in three commits: a synthetic fixture ledger and --demo flag for a stdlib HTML report generator; dark CSS theme and a summary strip in that single-file report; wrapping long lines in a CLI text card at 100 columns with tests, and rewriting a dataset card markdown	routine-coding	ledger-pinned	keppy	2026-W40
keppy:79c47a58	Add a --demo flag to a stdlib HTML report generator: bundled synthetic JSONL fixture with relative timestamps resolved at load, fixture directory, one new sidecar key, isolation tests	routine-coding	ledger-pinned	keppy	2026-W40
keppy:7499db80	Restyle a stdlib single-file HTML report: dark default theme from a palette dict, --theme flag, a four-cell summary strip computed from existing data, one new table column, tests for both themes	routine-coding	ledger-pinned	keppy	2026-W40
keppy:008a518e	Wrap long lines of a CLI text card at 100 columns with textwrap gated by a surface flag and a --wide escape hatch, keeping a byte-exact snapshot test for one surface; a seeded generator for a JSONL demo fixture; a today-count fix; rewrite a markdown dataset card builder and fix its stale paths	routine-coding	ledger-pinned	keppy	2026-W40
keppy:54c8387a	Add a CLI verb to a Python library that reads a JSONL ledger, selects rows by a method field, duplicates correction rows, merges optional JSONL tasksets and seed rows, dedupes on text, writes JSONL and prints a per-lane count summary with --json and --strict; tests on a fixture ledger	routine-coding	ledger-pinned	keppy	2026-W40
keppy:164ffcba	Add an optional classifier backend to a Python routing library: lazy-import a saved transformers model from a local directory, calibrated softmax, threshold abstain, wire into an existing classification cascade between rules and an LLM fallback, an install CLI verb that validates and copies the artifact, optional-extra...	routine-coding	ledger-pinned	keppy	2026-W40
keppy:4c3e6f45	Write an example training driver script for an existing in-process encoder fine-tune function: argparse with dry-run and backend flags, post-training evaluation on a held-out JSONL writing predictions and extra metrics keys including a confidence-threshold search, and a CPU test with a tiny random-init transformers mod...	routine-coding	ledger-pinned	keppy	2026-W40
keppy:be333918	Two Python scripts: one calls an OpenAI-compatible chat endpoint via urllib to generate paraphrased lines per category with dry-run and dedupe, the other does a seeded stratified train/eval split of JSONL rows with pair constraints and an audit JSON; tests with a monkeypatched client	routine-coding	ledger-pinned	keppy	2026-W40
keppy:2d619899	Scaffold a new Python repo for a JSONL dataset: pyproject, JSON schema, stdlib validator with regex publish gate and denylist, review script printing a batch sha, gated upload script tested with a fake client, fixtures and pytest, README; one commit	routine-coding	ledger-pinned	keppy	2026-W41
keppy:437a7d8c	Critique: 'A reward model trained to convergence on human preference comparisons has an optimum aligned with human values.' Give the two strongest failure modes, with a concrete mechanism for each.	alignment-reasoning	taskset	keppy	2026-W40
keppy:e4020717	Explain mechanistically why RLHF on thumbs-up/down feedback tends to produce sycophancy, and name two training-set or loss-design interventions that specifically target that mechanism (not just 'add diversity').	alignment-reasoning	taskset	keppy	2026-W40
keppy:ed47e644	A model is fine-tuned to maximize 'helpfulness' as scored by an LLM judge, and its measured helpfulness-by-human drops as judge-score rises. Name the phenomenon precisely, and explain the two conditions under the proxy regime that make divergence inevitable rather than accidental.	alignment-reasoning	taskset	keppy	2026-W40
keppy:523e23df	In circuit-level interpretability, what does it mean for a feature to be 'universally' shared across models vs 'model-specific', and why does that distinction matter for using interpretability results to certify safety of a NEW model?	alignment-reasoning	taskset	keppy	2026-W40
keppy:d9a4095e	State the scalable oversight problem in one paragraph, then evaluate debate (two models argue before a judge) as a solution: one argument for, one concrete failure mode with mechanism.	alignment-reasoning	taskset	keppy	2026-W40
keppy:63aba4f7	Contrast outcome-based RL (RLHF on a reward model) and process-based RL (rewards on process supervision / RRM). Name one concrete way each fails that the other resists, and one failure they share.	alignment-reasoning	taskset	keppy	2026-W40
keppy:b6a03099	The 'shards' framing says a value is a contextually-activated decision-making component shaped by reinforcement. Give one empirical prediction this framing makes that pure 'utility function over world-states' framing does not, and an experiment that could distinguish them.	alignment-reasoning	taskset	keppy	2026-W40
keppy:97ed281a	Why are conjunctive goals (many conditions must ALL hold) often argued to be safer for corrigibility than disjunctive ones (ANY of several paths to the goal works)? Give the mechanism, then name one way the conjunctive argument breaks down.	alignment-reasoning	taskset	keppy	2026-W40
keppy:5898a3fc	Distinguish (a) trained-in deception (gradient descent finds deceptive policies because they minimize training loss), (b) sandbox breakout from capability alone, and (c) misgeneralization that LOOKS like deception. For each, state one decisive test to tell them apart in a suspicious observation.	alignment-reasoning	taskset	keppy	2026-W40
keppy:422fde6b	Argue for or against: 'As models get more capable, external oversight (evals, red-teaming, monitors) becomes MORE valuable relative to interpretability, because interpretability matures more slowly than capability growth.' Steelman both sides, then state which empirical trend would settle it.	alignment-reasoning	taskset	keppy	2026-W40
keppy:424b5a0e	Write a numerically stable logsumexp(x, axis=-1) in NumPy. It must handle -inf entries and rows that are entirely -inf (return -inf, no NaN). Reply with one python code block.	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:2fcda0db	Implement GRPO-style advantage computation: given per-group rewards (list of lists), the advantage of each sample is (reward - group mean) / (group std + 1e-8) computed withing each group independently. Handle single-element groups (advantage 0). Reply with one python code block defining compute_advantages(rewards) -> ...	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:ed3f23ce	Implement a cosine LR schedule with linear warmup: lr(step) = base_lr * (step/warmup) for step < warmup, else base_lr * 0.5 * (1 + cos(pi * (step - warmup) / (total - warmup))). Handle step=0 and warmup=0 (skip warmup phase). Reply with one python code block defining lr_at(step, base_lr, warmup, total).	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:eb50099f	Given a list of per-microbatch loss gradients (each a list of floats, one per parameter), implement gradient accumulation with scaling: accumulated = sum(gradient_i / n_microbatches) per parameter, then divided by 1.0 (identity). No numpy needed. Reply with one python code block defining accumulate(grads) -> list of fl...	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:3ae8ff5d	Implement a minimal BPE tokenizer trainer: given a corpus string and a number of merges n, repeatedly merge the most frequent adjacent pair of characters until n merges are done or no pair repeats. Return the merge list (pair tuples, in order). Only characters, no pre-tokenization. Reply with one python code block defi...	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:81c8f2db	Implement dynamic loss scaling for mixed-precision training: given a list of grad_norms (one per step) and the initial scale, return for each step a tuple (skip_update, scale_used). skip is True when the grad_norm is inf or nan. The scale doubles after every non-overflow step and halves (minimum 1.0) after an overflow ...	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:b5aeb5cb	Implement an epoch boundary helper: given dataset length n, batch size b, and step t (0-indexed, counting optimizer steps), return (epoch, step_in_epoch). The last batch may be smaller than b; an epoch ends when all n samples are consumed. Reply with one python code block defining epoch_of(n, b, t).	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:a0ce5e1c	Implement early stopping: given a list of validation losses (one per epoch) and patience p, return the epoch index at which training should stop, i.e. the first epoch after which p consecutive epochs showed no improvement over the best loss so far. Equal loss is not an improvement. Reply with one python code block defi...	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:00530b0d	Implement decoupled weight decay (AdamW) update for one parameter: given current weight w, gradient g, m, v (EMA states), hyperparams lr, beta1, beta2, eps, wd, and timestep t, return the new weight (bias correction applied to m and v; decay applied to w directly, not the update). No numpy. Reply with one python code b...	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:af68e156	Implement topological sort of a DAG of computational-graph nodes: given edges as (from, to) pairs and a list of all node ids, return one valid topological order (list). Raise ValueError on a cycle. Reply with one python code block defining topo_sort(nodes, edges).	dl-ml-research-engineering	taskset	keppy	2026-W40
keppy:a13a4847	Write a slugify(text) function: lowercase, replace any run of non-alphanumeric chars with a single hyphen, strip leading/trailing hyphens, collapse repeated hyphens. Empty result -> empty string. Reply with one python code block.	routine-coding	taskset	keppy	2026-W40
keppy:5cc02611	Write parse_log(line) that parses an Apache combined-format log line: 'IP - - [date] "METHOD PATH PROTO" status bytes'. Return a dict with keys ip, date (the bracketed string, no brackets), method, path, status (int), bytes (int). Return None if the line doesn't match. Use a single regex. Reply with one python code blo...	routine-coding	taskset	keppy	2026-W40
keppy:5263d1ca	Write count_by(csv_text, column) that takes CSV text with a header row and a column name, and returns a dict mapping each distinct value in that column to its row count (excluding the header). Use only the csv module or plain split; handle quoted fields by using the csv module. Reply with one python code block.	routine-coding	taskset	keppy	2026-W40
keppy:36c81b40	Write render(template, values) that substitutes '{{key}}' placeholders in a string from a dict; unknown keys render as empty string; '{{' or '}}' without a key inside are left as-is only when malformed (e.g. '{{}' stays '{{}'). Values are str()'d. Reply with one python code block.	routine-coding	taskset	keppy	2026-W40
keppy:02d7011b	Write backoff(n, base=1.0, cap=60.0) returning the list of delays for retry i in 0..n-1: delay = min(base * 2**i, cap), computed for i = 0,1,...,n-1. Reply with one python code block.	routine-coding	taskset	keppy	2026-W40
keppy:674cc6bc	Write to_roman(n) converting 1..3999 to a Roman numeral string. Raise ValueError outside that range. Reply with one python code block.	routine-coding	taskset	keppy	2026-W40
keppy:1146a46e	Write flatten(d) that flattens nested dicts into one level with dotted keys: {'a': {'b': 1, 'c': {'d': 2}}, 'e': 3} -> {'a.b': 1, 'a.c.d': 2, 'e': 3}. Non-dict values are leaves. Empty dicts disappear. Reply with one python code block.	routine-coding	taskset	keppy	2026-W40
keppy:1250673b	Write a pure function bucket(events, capacity, refill_per_sec) simulating a token bucket: start full. Each event is a timestamp in seconds (ascending, first may be 0). Before each event, refill by elapsed-since-previous-event * refill_per_sec (capped at capacity; the clock starts at 0). The event is allowed only if at ...	routine-coding	taskset	keppy	2026-W40
keppy:6a491bc6	Write week_starts(start, n) that given an ISO date string 'YYYY-MM-DD' returns a list of the n ISO dates of the Mondays of consecutive weeks starting with the week of the given date (Monday of that week first). Use only the datetime stdlib module. Reply with one python code block.	routine-coding	taskset	keppy	2026-W40
keppy:a313021d	Write apply_edits(lines, edits) applying a simplified edit script to a list of strings. Each edit is a tuple: ('ins', line_no, text) inserts text BEFORE the 1-indexed line_no, or ('del', line_no) deletes that line (2-tuple, no text). Edits arrive in DESCENDING line order so earlier edits don't shift later targets; appl...	routine-coding	taskset	keppy	2026-W40

evalroute-tasks

A small human-labeled corpus for training and evaluating evalroute's task→lane router. Each row is one real coding/research/writing task with the lane a human asserts it belongs to. Tier 1 (task → lane) is the only required label today; the schema leaves room for tier 2 (task → arm outcome) and tier 3 (task → done artifact) so contributors can backfill later without a format break.

Status: zero to one. One contributor, 54 rows, four of nine lanes. Per lane: routine-coding 30, dl-ml-research-engineering 13, alignment-reasoning 10, prose 1; hard-agentic-coding, long-doc-reading, web-research, math-first-principles and orchestration have none yet. 24 rows are real tasks pinned in one dispatcher's ledger (mostly coding briefs), 30 are hand-written tier-a taskset rows. n is small and the mix reflects one person's work — that is the point of publishing it: your work shapes your router, and the shared encoder is only as broad as this file. The trained encoder is keppy/evalroute-lane-encoder; the outcomes flywheel (lane/arm/verdict, no text) is keppy/evalroute-flywheel.

Tiers

Tier	Question	Keys	Required
1	which lane?	id, text, label, source, contributor, week	yes
2	which arm, and did it work?	arm, verdict, method, n	optional (all-or-none)
3	how was done judged?	checker / rubric / reference (at most one)	optional

Row example

{"id": "keppy:a3f9c1", "text": "add a CLI verb that reads the ledger and prints per-lane counts",
 "label": "routine-coding", "source": "ledger-pinned", "contributor": "keppy", "week": "2026-W40",
 "arm": "z-ai/glm-5.3-flash@medium", "verdict": "pass", "method": "measured", "n": 10,
 "checker": null, "rubric": null, "reference": null}

The publish gate (three steps, all must pass)

Mechanical validation: python scripts/validate.py tasks/*.jsonl checks id/text/lane/source/contributor/week format, tier-2 and tier-3 coherence, and a secret/PII/hostname regex gate over text and tier-3 strings.
Denylist: if private_terms.txt exists in the repo root (git-ignored) or you pass --private-terms <file>, any case-insensitive whole-word hit fails the row. Keep employer/project terms you must not publish in that file.
Human review: run python scripts/review.py tasks/<you>.jsonl, read every numbered row, compute the batch sha, add reviewed: <sha> <date> <contributor> to REVIEWED.md, then open a PR. scripts/upload.py refuses to touch Hugging Face unless REVIEWED.md contains a line for every file's sha.

How to contribute

Fork/branch, add your own file tasks/<your-handle>.jsonl (one JSON row per line; the file stem and row-id prefix must be your handle).
Run python scripts/validate.py tasks/<your-handle>.jsonl — it must exit 0.
Run python scripts/review.py tasks/<your-handle>.jsonl, actually read the rows, and add the printed reviewed: line to REVIEWED.md.
Open a PR. A reviewer re-runs validate and diff-checks your rows.

What is not here

No paraphrases: train F showed paraphrases that outnumber real rows pull the classifier toward the paraphrase dialect. No ledger exports without review — a row lands here only after a human decided the text is safe to publish. No personal or employer details: names, internal hostnames, credentials, and project-denialist terms are caught by the gate, but judgment is still yours.

License

Downloads last month: 28

Number of rows:

Total file size:

28.3 kB

Models trained or fine-tuned on keppy/evalroute-tasks

keppy/evalroute-lane-encoder

Text Classification • 37.2M • Updated 3 days ago • 27