Hugging Face's logo
Murtadha hamid's picture

Murtadha hamid PRO

murtadha9
1 4
luke1996's profile picture webbrain-one-850928's profile picture
Ā·

AI & ML interests

Building applied AI systems: models, data pipelines, and monitoring & analysis tooling across text, speech, and vision ā€” with a focus on Arabic and the Iraqi dialect

Recent Activity

reacted to sergiopaniego's post with šŸ”„ 1 day ago
if you want to deep dive into how the new @MistralAI Large 4 was built, we sat down with the team to learn straight from them! scaling, specialization, evals, cyber, sandboxing and why open weights matter a really detailed look at what it takes to build and release an open frontier model, super recommended https://www.youtube.com/watch?v=FczbSauVvjw
reacted to theirpost with šŸ”„ 2 days ago
Glyph Lens Instead of retyping printed text from a photo or scanned page word by word, you get it back in Arabic or English, ready to copy or download. It also highlights the words it may have misread right on the image, so you only need to check those. Free, runs in the browser, stores nothing. šŸ”— https://huggingface.co/spaces/murtadha9/glyph-lens
reacted to savoji's post with šŸ”„ 2 days ago
One Qwen3-VL-8B checkpoint improves all three streaming spatial benchmarks at once: OST-Bench 50.52 to 63.93 (+13.41), UCS-Bench +7.44, OVO-S-Bench +4.91. The change is in the supervision, with no new memory module or annotation. The ChronoReplay paper calls the failure mode interface lock-in: post-train on one answer format and that readout improves while others stall. A categorical-only control hits 58.03 on OVO-S but drops OST from 50.52 to 47.48. šŸ” The recipe ā€¢ Keep targets in native formats (option, metric value with units, timestamp) ā€¢ Replay episodes in source order, each target attached to its causal prefix ā€¢ Loss mass: 2/3 on 33,180 single-turn examples, 1/3 on 6,381 episodes (alpha = 0.5) šŸ“Š The numbers ā€¢ Three-benchmark mean 56.86, against 50.74 (categorical-only) and 50.76 (trajectory-only) ā€¢ Chronological order beats a shifted-order control by about 2.41 OST points on two seeds; the 13.41 belongs to the whole recipe ā€¢ Adaptive frame selection adds +4.09 on OVO-S (57.98) with 80.50 frames instead of 128 āš ļø Caveats ā€¢ One backbone, indoor scenes only; the single-family comparison is confounded by data coverage ā€¢ The repo has a CPU smoke test and a --dry-run training command, but no license and no fine-tuned weights https://huggingface.co/papers/2610.34283 https://arxiv.org/abs/2610.34283 https://github.com/yihaohu0118/ChronoReplay If you fine-tune VLMs for embodied work, do you score categorical, metric and temporal readouts together? šŸ“š Read it elsewhere ā€¢ The deep dive on Medium: https://medium.com/p/8d2fb1bf832d ā€¢ The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/chronoreplay-lifts-qwen3-vl-8b-by ā€¢ The thread on X: https://x.com/saavoji/status/2116229682844877156 ā€¢ Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#chronoreplay
View all activity

Organizations

None yet