Dushyant Rajput
rajpdus
·
AI & ML interests
None yet
Recent Activity
updated a model 26 days ago
altslate/JugnuLM-110M-R4b updated a model 26 days ago
altslate/JugnuLM-110M-R4a updated a model 26 days ago
altslate/JugnuLM-110M-R3Organizations
Can we have the text for the dataset?
#1 opened about 2 months ago
by
rajpdus
Value residual scaling, the 25.8% vocabulary tax, and horizontal state grounding in deep-thin SLMs
2
#1 opened about 2 months ago
by
AndrewThompson1233
Add JugnuLM-110M-R2+ (altslate) Sub-150M, pretrained from scratch (Qwen3-arch + value residuals + Muon, WSD schedule, ~25B tokens on FineWeb-Edu). Open training checkpoint (not merged): https://huggingface.co/altslate/JugnuLM-110M-R2plus Zero-shot, LM-eval harness (acc_norm), + AxiomicLabs ArithMark-3 script: arc (ARC-Easy) 47.64 | hellaswag 36.74 | piqa 65.29 | arcChall (ARC-Challenge) 26.96 | arithmark-3 38.60 109,737,302 params. (arithmark-2 omitted — dataset not publicly available.) NOTE for verification: this is a value-residual model, so it must be loaded with trust_remote_code=True (the repo ships modeling_jugnu_vr.py + auto_map). Loading it as a stock Qwen3 silently drops the value-residual pathway and understates the scores (~6 pts ARC-Easy, etc.). The ArithMark-3 script loads it correctly by default.
1
#154 opened about 2 months ago
by
rajpdus
Add JugnuLM-110M-R2+ (altslate) Sub-150M entry: JugnuLM-110M-R2+ (109.7M params) — value residuals + Muon, WSD schedule, 25B tokens. Self-reported (lm-eval-harness, VR-aware load): BLiMP 82.52, ARC-Easy 55.13, WikiText-2 byte-ppl 1.8735. Model: https://huggingface.co/altslate/JugnuLM-110M-R2plus (load with trust_remote_code=True — value-residual pathway; stock from_pretrained silently degrades it). Recipe: https://github.com/AltSlate-Labs/jugnu
🚀 1
#71 opened about 2 months ago
by
rajpdus
Add JugnuLM-53M (altslate) to the leaderboard
#1 opened 2 months ago
by
rajpdus
Add JugnuLM-53M (altslate) to the leaderboard
#67 opened 2 months ago
by
rajpdus