Downloads · 30 days
0
anlp-a2-p1-moe_iso_active
siddarthg44/anlp-a2-p1-moe_iso_active
anlp-a2-p1-moe_iso_active is a translation model from siddarthg44. Use it when you need text moved from one language to another. It is set up for pytorch.
Part 1 of ANLP Assignment 2: a decoder-only transformer whose feed-forward layers use the moeisoactive variant, trained on belumind/en-vi-ja-curated-500k-triplets to translate Vietnamese and Japanese into English.
Access
Public
Updated Sep 29, 2026
Repo size
602 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt201 MB · 99%
At a glance
- Task
- Translation
- Library
- pytorch
- Access
- Public
- Created
- Sep 11, 2026
- Updated
- Sep 29, 2026
- SHA
- b9bc7e21
Try a prompt
Model card
From the Hugging Face model README
siddarthg44/anlp-a2-p1-moe_iso_active
Part 1 of ANLP Assignment 2: a decoder-only transformer whose feed-forward layers use the
moe_iso_active variant, trained on belumind/en-vi-ja-curated-500k-triplets to translate Vietnamese and Japanese
into English.
| FFN variant | moe_iso_active |
| FFN config | {'type': 'moe', 'd_ff': 2048, 'n_experts': 4, 'top_k': 2, 'n_shared': 0, 'match': 'active'} |
| d_model / layers / heads | 512 / 8 / 8 |
| Context | 256 |
| Token budget | 150,000,000 |
| Tokenizer | byte-level BPE, 16,000 merges, trained jointly on en+vi+ja |
Files: best.pt (weights + config), tokenizer.json, eval.json (test perplexity, BLEU and
expert-specialisation metrics).
Load with src/model.py:
state = torch.load("best.pt", map_location="cpu", weights_only=False)
model = build_model(TransformerConfig.from_dict(state["config"]))
model.load_state_dict(state["model"])
- Task
- Translation
- Library
- pytorch
- Languages
- en, vi, ja
- Created
- Sep 11, 2026
- Updated
- Sep 29, 2026