Hodge-3.5
Hodge-3.5 is a ternarized causal language model. The linear layers are Hodge linear, quantized offline.
Model
| hidden size | 2560 |
| layers | 30 |
| intermediate size | 6912 |
| attention heads | 20 |
| key-value heads | 5 |
| head size | 128 |
| vocabulary | 514304 |
| context length | 4096 |
| rope theta | 500000 |
| activation | relu2 |
| embeddings | tied |
| weight dtype | bfloat16 |
| bos token id | 128000 |
| eos token ids | 128001, 128009 |
Weights
hodge-3.5.bnz
Chat
Each turn is a role line closed by <|eot_id|>. Generation starts at Assistant: .
Sampling
generation_config.json samples with temperature 0.6 and top-p 0.9. The length cap is 4096.
Run
ModelRun loads this directory. The catalog points model at hodge-3.5.bnz, beside config.json and tokenizer.json.
modelrun run --config hodge-3.5.bnz.yaml