Spaces:
Runtime error
Runtime error
Commit History
-
Claim Tiny Titan badge: add achievement:tiny-titan Space tag
The write-up and README already claim Tiny Titan (Qwen2.5-VL-3B, under the 4B line), but the Space metadata only declared offgrid/llama/fieldnotes. Add the achievement:tiny-titan tag so the badge is registered on the official hackathon Space. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>68572d7 -
Minor edits
Verified signatureNew thumbnail, off-brand achievement updates.f638365 -
README/write-up: surface demo CTA, claim Tiny Titan, link field notes
Address the friendly-review feedback: promote the live-demo link to a prominent CTA under the title, foreground the Tiny Titan (<=4B) claim in the badge line and merit-badges section, and add a closing field-notes pointer. Launch-post link confirmed correct (review's "broken" flag was an X bot-wall false positive). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>b5dee12 -
Fold doc corrections through writeup + README prose
- writeup: second 'short ready tasks' instance (line 29) -> 'short, high-impact' (matches QuickWin = impact+effort, not readiness). - writeup + README 'How it works': prompts.py asks for JSON / llm.py grammar- constrains it (json_object), reflecting the prefill removal in f92e318. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>d11afff -
Field Notes link + doc factuality fixes (QC vs code truth)
- README: link the honest write-up (backs achievement:fieldnotes). - prompts.py: docstring described the removed JSON-prefill trick; the code uses response_format json_object grammar (since f92e318). Corrected to match llm.py. - README: first-boot download is ~3.3 GB (1.9 GB model + 1.3 GB f16 mmproj), not ~2.3 GB — the projector was uncounted. - README + writeup: QuickWin keys on impact+effort, not readiness; 'short, ready' -> 'short, high-impact' (QW = 1 + 0.6*(I/10)*exp(-E/1.5) in score.py). - README: drop stale HTML comment claiming a nonexistent front-matter video. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>744013f -
Claim Tiny Titan: declare <=4B model + foreground the small-model story
Tiny Titan is a judged bonus (no self-declare tag). Add a models: frontmatter entry (ggml-org/Qwen2.5-VL-3B-Instruct-GGUF) so the <=4B qualification is machine-readable, and surface '3B params (<=4B)' + offline/llama.cpp in the header per the evaluator's 'make the small-model story loud' guidance. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>38ad4ca -
LICENSE: fill in copyright line
Replace the Apache appendix template with "Copyright 2026 OptiQuant Analytics". Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>78a0e3d -
Add required social-media post link to README
Links the X launch post per the Build Small submission requirement (social post showcasing the app, linked from the Space README). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>6044f90 -
Add Apache-2.0 LICENSE file
Matches the license already declared in the README and HF Space front-matter. Canonical text so GitHub detects it and shows the badge. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>b9d621d -
Add hackathon submission tags: Backyard AI track + offgrid/llama/fieldnotes badges
Canonical namespaced slugs read by the official ready-to-submit evaluator (checks.py): track:backyard plus achievement:offgrid/llama/fieldnotes. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>0ad5433 -
README: render cleanly on GitHub and Hugging Face
GitHub strips the HTML <video> tag, so the demo showed nothing on the repo page. Replace it with a clickable poster image (renders on both GitHub and HF) plus explicit text links to the video and live Space. Add a links/badges line under the tagline. HF front-matter is untouched. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>5ca0db4 -
Restyle hero: brand-gold accents on darker ink->indigo->blue gradient
Match the demo video / what-first.com hero: accent the word 'small' and key phrases in GOLD_LT, darken the gradient, and force white text over Gradio's prose defaults. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>0495e5e -
Merge Space deploy history (866bea8): same speed+UX changes, already present locally
The Space's 866bea8 made identical edits to app.py/llm.py/prompts.py/score.py/ start.sh/test_score.py as local 6408b42; local additionally carries the demo/ video + writeup. Merging reunites the two published histories so GitHub and the HF Space share one lineage. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>65768a2 -
demo video and writeup
3bb9ac2
-
Speed + UX: trim scored schema, cache prompt, lighter ctx, live status
App-only deploy to the Space. Same app changes as origin commit 6408b42 but without the demo/ video tooling (and its font binary, which HF rejects and the running app never uses). - PARSE schema trimmed to ranker-only fields (~20-35% fewer output tokens) - Why column derived from score.explain (deterministic, zero tokens) - cache_prompt: reuse system-prompt KV between calls - llama-server ctx 8192 -> 4096 - prioritize() streams an instant status so CPU inference isn't a frozen wait Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>de744cc -
updates to speed up computation of ranked list
a56a725 -
Fix JSON coercion: every LLM call returned empty (ported prefill broke)
The cloud app prefills the assistant turn with '{' and lets Claude continue it; the Anthropic API honors that natively. llama.cpp's chat endpoint does not continue a trailing assistant turn — it generates a fresh complete object starting with '{', so prepending '{' produced a doubled brace ('{{ ... }'). _extract_json never reached depth 0 and returned {}, so parse/extract/score all yielded zero items: "Nothing actionable found" on every input, and a silent no-op on "Re-suggest with AI". - Drop the prefill; use llama.cpp response_format json_object, which forces a single valid JSON object as a sampling grammar (no prose, no fences, no unbalanced output) — the reliable equivalent for a local 3B model. - Make _extract_json recover complete items from a truncated array so one cut-off task can't void the whole batch (also defends against the old double-brace shape). Verified against clean/double-brace/truncated/prose. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic…dc45861 -
Downscale screenshots before the vision encoder to cut CPU latency
Qwen2.5-VL's dynamic resolution turns a full-size screenshot into thousands of vision tokens, which dominates inference time on cpu-basic. Cap the longest edge at 1280px (env-tunable) before encoding — keeps a task list readable while cutting vision-token count and prefill several-fold. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>2738645 -
Cap llama.cpp build parallelism to avoid OOM on HF Spaces builder
Bare `-j` used all cores; parallel g++ on ggml kernels blew past the cpu-basic build RAM limit (exit 137 OOMKilled). `-j 2` keeps it bounded. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>a14cb58 -
Loosen gradio pin to a safe 5.x range for reliable Space builds
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>66d6761
-
Trim design rationale from score.py; tighten README lineage
Keep the scoring code identical (tests unchanged, 16/16) but cut the verbose "why" comments that doubled as a design doc. Correct the README: in the product, the cloud model proposes task fields and a deterministic engine ranks — the small local model here replaces the language work. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>dfbc644
-
Pin LF line endings for Linux container
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>caff696 -
whatfirst-small: offline Gradio + llama.cpp task prioritizer
HF Build Small hackathon entry (Backyard AI). A small local VLM (Qwen2.5-VL-3B via llama.cpp) turns a brain-dump or a photo of a to-do list into structured tasks; a clean-room Python port of the what-first.com scoring engine ranks them transparently. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>3a8d5b6