-
v5: ONNX decoders for the browser runner
-
v5 is the GRPO model, v5-base the one before it (v5-rl kept as a copy)
-
Upload v5/tokenizer.json with huggingface_hub
-
Upload v5/ckpt_final.pt with huggingface_hub
-
Upload v5-rl/tokenizer.json with huggingface_hub
-
Upload v5-rl/ckpt_final.pt with huggingface_hub
-
Upload v4-large/tokenizer.json with huggingface_hub
-
Upload v4-large/ckpt_final.pt with huggingface_hub
-
v4: v4/tokenizer.json
-
v4: v4/ckpt_final.pt
-
v3 tokenizer
-
v3: 113.8M, 24B tokens on corpus_full (7 sources, strict dedup, EOS docs); val 0.7495; serve at t=1.1 top_k=50
-
Upload v2-100m/ckpt_final.pt with huggingface_hub
-
Upload v2-100m/tokenizer.json with huggingface_hub
-
Upload v2-100m/ckpt_final.pt with huggingface_hub
-
Add v2-production tokenizer
-
Add v2-production checkpoint (25M, LAMD-augmented corpus, loss 0.93)
-
Add v2-pilot tokenizer
-
Add v2-pilot checkpoint
-
Add model card
-
initial commit