H
Hi tbo81, The two links you shared don’t seem to be working. When I clicked them, both redirected me to a 404 error page. Could you please share the correct links? I’ll review them and provide you with the requested details. Thank you!
H
Both links 404 as written — the files are in the repo as RoadMAP.md and Evolution_LOG.md, so renaming the files to the names you linked is the fix that makes every copy of those links work. On the roadmap itself: it lists the ablation as proven — neither damp alone nor inflated v alone beats Adam, both together do — but the evolution log has no entry for that run, and your post #9 here had the split as the next thing to run; if it has run, it needs its log entry with seeds and numbers, and if…
H
We published two new files for Peanut: Roadmap: https://github.com/dskc/peanut/blob/main/ROADMAP.md - where we are going, 6 baselines, 7 metrics, 4 unproven gaps to close Evolution Log: https://github.com/dskc/peanut/blob/main/EVOLUTION_LOG.md - dated log of each experiment with results, like release notes, including failures Our direction: CPU-only is a design decision, not a shortcoming. We will prove Peanut with same datasets, same models, properly tuned LRs, 5 prime seeds. Tell us: Are we…
H
Follow-up to #32: the simpler way to apply it A rule is written as one sentence in the tool’s description, with a label. Before the tool runs, code reads that description itself and counts what is not confirmed. Nothing is coded per rule. [Required] The sale must be agreed, as shown by the order status. [Verify] The time is the one the seller states. It is one moment. [Notice] The buyer sees the time and location immediately. [Required]: confirmed only from system data. The user’s words do not…
H
The second in a series: we run RoboAudit on public LeRobot datasets and publish the report cards, including the parts where the tool itself is wrong — one card in this issue is exactly that. Scores are relative indicators, not absolute quality measures — and nothing here says anything about how well a policy trained on the kept episodes will perform. Tool: RoboAudit 0.3.0 · 12 rule/threshold checks · full audit (no sampling). The xarm card was re-audited with RoboAudit 0.3.1 after a reader bug…
H
Since you asked for failure cases, a few patterns from Hebrew legal text worth adding to the eval set. Rulings that already pseudonymize parties as פלוני / אלמונית (those shouldn’t count as misses), titles directly before the name like עו"ד or כב’ השופט, and Teudat Zehut numbers, which are 9 digits with a check digit you can validate to tell them apart from case or file numbers. Text that went through OCR is a different distribution from typed text: final letters (ם/ס, ן/ו) get confused and…
H
I’ve been developing Model-Native Structural Representation (MNSR), a research framework investigating how the representation of task structure influences large language model interpretation, reasoning, and execution. The central premise is that task structure is not equivalent to its token serialization. Tasks contain consequential relationships—including scope, identity, precedence, dependencies, constraints, state, and verification—that must be reconstructed from the model-facing…
H
Repository ArchitecturalGravity/ArchitecturalGravity GitHub - ArchitecturalGravity/ArchitecturalGravity: A cross-model Agent Control Language for... A cross-model Agent Control Language for structuring, composing, and operating on context directly within natural-language prompts, enabling more precise reasoning, consistent execution, and efficient token usage while remaining fully prompt-native, infrastructure-free, and portable across models, interfaces, agents, and evolving workflows…
H
Thank you, Lawrence, and Cael. It’s encouraging that a judge given the rule in words could tell whose money it was, and I appreciate how clear you were about what this doesn’t show. What stands out to me is that spelling the principle out as ownership and consent is what made it testable. I’d love to see how it does on harder, messier cases someday.
H
The scope key point is the one that decides whether this check is usable. If the result carries its own declared scope rather than inheriting one from wherever the reader happens to find it, then a child’s call-17 can never be resolved against the parent by accident. I would also keep ambiguous declarations as a distinct outcome rather than folding it into a failure count, since a retry and an amendment look identical until you have an attempt field. Adding attempt identity alongside the scope…
H
And it also needs to take into account iMatrix. Because on Apple Metal, I’m finding running iMatrix on standard K quant like Q6_K and Q4_K_M are making quite a difference over the standard IQ file naming IQ people are used to using. Real difference in prefill and decode and in KLD. I run llama.cpp and Asher/gmlx (gguf on mlx vm) as my main runnners on M1 Max (32 core gpu I inspect every single GGUF I download for what the layer layouts are because those matter to me on metal because it is not…
H
@capitaharlock Yes - holding the emissions separately makes this a much stronger test. A reopened ticket is not evidence that no customer email escaped. I’d pin down the release point: the version check and authorization to release the outbox entries need to share an enforced ordering boundary with incoming messages. If release is a later worker step, a revocation ordered before release must invalidate those pending entries, including delayed retries; otherwise the same race just moves into the…
H
Thank you — for the original suggestion, and now for this one. Manufacturing the missing class instead of waiting for a capture-time dataset was the right call, and we ran it the same day. What we did. From the vials dataset’s 60 clean episodes we took 12 with strong mid-motion travel, deleted runs of 2, 3 and 5 frames at the motion, and kept 48 untouched episodes as controls — in two variants: (a) frame_index re-sequentialised, timestamps original; (b) frame_index left gapped. Then a full…
H
@WilliamFrye73 No, that part isn’t from your analysis. It’s how chat templates usually work. Qwen3’s default template drops the content of earlier, finished turns when it rebuilds the history for the next prompt, so the model may never have seen that block again. If your logs kept the raw prompt sent for the turn after the incident, search it for the block. If it isn’t there, the model had nothing to react to, and its silence says nothing about concealment.
H
You may not need to wait for a capture-time dataset to validate the split — both classes can be manufactured from episodes you already have. Take smooth episodes with no A2 findings, delete a short run of frames mid-motion, and keep each surviving frame’s original index as its timestamp: the interval ratio gets a known gap to find, and motion that was ramping across the deleted frames should arrive as one simultaneous state jump with no ramp. Set those beside the 15 real findings — state…
H
nootxlm: The published build is 18,725,978,368 bytes — roughly 3.2 times smaller. Following up with the companion testing from today, in the same spirit of measuring before claiming: Diffusers: Diffusers #15305 — tested the fix in PR #15307: exact Accelerate: Accelerate 1.15.0 — warning_once keeps adapter instances alive (measured) — @nootxlm
H
Hello — sharing a build we published on the Hub: nootxlm/Llamatron. It starts from NVIDIA’s Nemotron-3-Nano-30B-A3B checkpoint, about 59 GB in its stock form. The published build is 18,725,978,368 bytes — roughly 3.2 times smaller. The shrink comes from importance-guided rounding: the weights that matter least are rounded hardest, and the ones that carry the model are kept fine. We checked quality against plainer copies of the same shrink as we went; the guided build held where the plain ones…
H
@wihaha to the task ID rather than the credential. A task that resumes after a disconnect may come back with a fresh credential, and that shouldn’t reset the count. Binding it to the task ID only helps if the agent can’t touch the reset either. Let the orchestrator clear the count at a task boundary and nothing else, so a stretch of good behavior never earns the budget back.
H
@nootxlm, here is the result of the plan from my last post, as promised. Setup (fixed before the GPU run): I rewrote 33 benchmark traps so that only the feature the catching rule keys on changes, and paired each with a twin control fact from the same domain. τ was tuned only on items where R1–R7 all stay silent, over a grid from 0.30 to 0.95. The objective was the mean of the trap catch rate and the control pass rate. The chosen τ and the default were then evaluated exactly once on 88 untouched…
H
nootxlm: PR #15307, at its current head: the converted delta matches exactly — max difference 0.0 — and the converted lora_A equals the original. Keeping the result where it belongs: exact at the current head, and lora_A untouched. The companion finding from the same day’s testing is here: Accelerate 1.15.0 — warning_once keeps adapter instances alive (measured) — @nootxlm
H
@SheilaMayoAI on what the close triggers, I’d split it: the status flip is the only reversible part, and the emissions (the customer email, the CSAT survey) sit in an outbox that only releases after the final version-checked commit. Then the compensation test can inject the revoking message between close and reopen and assert on what actually went out, not on the ticket status, which will look fine after the reopen.
H
Hello — a small memory finding on accelerate, shared here since we can’t open it on GitHub right now. We couldn’t find an existing report of it. In accelerate 1.15.0, MultiProcessAdapter.warning_once is decorated with lru_cache (accelerate/logging.py, line 82), so the cache key holds self, and an adapter instance can’t be freed after its first warning_once call. Measured on 1.15.0 (Python 3.12, PartialState initialized): a fresh MultiProcessAdapter is freed on del + gc.collect() before any call…
H
Hello — a short test report on the LoRA converter fix, shared here because we can’t comment on GitHub at the moment. MaiAditya’s pull request #15307 swaps only lora_B in the final-layer remaps. We ran the issue’s own reproduction against its head today (torch 2.15.0, CPU): Stock diffusers 0.43.0: the converted delta does not match — max difference 39.0527, the reporter’s value. PR #15307, at its current head: the converted delta matches exactly — max difference 0.0 — and the converted lora_A…
H
Note: The Show and Tell category is for sharing and discussing projects, showcasing your Spaces, Models, Datasets and more. We value open-source and technical details over promotional content, so focus Hey everyone, Out of curiosity, I started exploring how efficiently language models could run on ordinary CPUs. I ended up building Sprout, a ~2 MB inference engine written in C, with GGUF support, AVX2 and AVX-512 optimizations, and built-in web research. I also developed a programming language…
H
I released SwarmLights-1K v2.1, a synthetic benchmark for tracking every drone in a dense night-time drone light show from three calibrated cameras. The odd result: with three cameras and exact calibration, my 3-D trackers still score below the best single camera on all 8 shows. Best single camera: IDF1 0.494–0.703 (2-D, pixels). Best three-camera baseline: IDF1 0.325–0.525 (3-D, metres). The two scores aren’t strictly comparable, but the gap is large and consistent. With exact detections, the…
H
If a ZeroGPU pilot would help and you can’t get a free slot, the Hugging Face Spaces grants guide describes opening a Discussion on a public Static Space to request ZeroGPU. Company is one of the project categories listed. For the dedicated 24 GB run, I’d share the pilot’s peak VRAM and runtime with the Spaces team and ask which grant route fits.
H
Tested your interval idea — here is what the data actually said We ran the test immediately, on all 15 A2 findings in the vials dataset. Short version: the test cannot speak on this dataset, but checking why led somewhere better. 1. The interval test is blind here. Every ratio is exactly 1.000. This dataset’s timestamps are frame_index ÷ fps by construction — all 18,175 intervals across 75 episodes sit exactly on the 1/30 s grid. The column holds no capture-timing information, so it cannot…
H
Thank you for the numbers, and for the correction — that changes what your first post was. A shift from 0% to below −89% in five components is not a warning about test design anymore; under the harsh profile, load did not shade the ranking, it broke those five outright. Two small things the numbers open up. The exact test on the data you already have costs no new experiment — the arithmetic rerun alone could settle whether the five clear the corrected level. And if you can name the five…
H
Thank you, it really is a sensible approach especially for a beginner. I have decided not to buy an expensive board to accommodate both cards but to invest in 64gb ram instead to run Strata with a single 7900xtx.
H
Thanks, Haifeng. Separately annotating the evidence about coverage and admission looks useful as a diagnostic. I would keep uncertainty in those intermediate annotations rather than force both into booleans: an unestablished coverage guarantee is not the same as established incompleteness, and an entry that cannot be inspected is not the same as an inspected record with no entry. A positive entry also needs to establish admission for the relevant task and current cycle. A stale entry is not…