Modujo-AI's picture
Add Modujo 4B A0.66B PT weights and loading instructions
510566b verified
|
Raw History Blame Contribute Delete
2.27 kB
---
library_name: transformers
pipeline_tag: text-generation
tags:
- pretrained
- pt
- mixture-of-experts
- qwen4-exp
language:
- zh
- en
---
# Modujo 4B A0.66B — PT
Approximately 3.96B total and 0.66B active parameters. This is an intermediate **continued-pretraining (PT)** snapshot, not an instruction-tuned assistant and not a completed 100B-token run.
## Architecture and lineage
28 layers, hidden size 1024, 32 routed experts, top-2,shared expert,tied embeddings. One pass, no trained QSA indexer. Linear-attention and dense-attention layers. Configured maximum positions 32K; current continuation training uses sequences up to 8192.
The snapshot comes from the 4B continuation segment on the pt_100b_ecommerce_20261005 corpus. Snapshot provenance and progress are in release.json. The corpus/run's 100B target is a plan, not a claim of tokens consumed by this snapshot. Tensors are stored FP32; FSDP model shards were merged for this release. Training optimizer/RNG files are not included.
The SFT directory in this repository was initialized from an earlier PT snapshot; it is not fine-tuned from this later PT release. Do not treat PT and SFT as an equal-step controlled comparison.
## Loading and completion
Download the subdirectory:
```python
from huggingface_hub import snapshot_download
snapshot_download('Modujo-AI/ModujoMoE',
allow_patterns=['pt/Modujo-4B-A0.66B/*'],local_dir='./modujo-download')
```
Run in the compatible environment recorded in tested-environment.json. The environment must provide qwen4_exp; the included compatibility helper disables an absent QSA indexer while preserving causal attention.
```bash
python modujo-download/pt/Modujo-4B-A0.66B/inference.py --device cuda:0 --raw --prompt '中国的首都是' --max-new-tokens 32
```
Use raw text completion for base-model inspection. The accompanying native chat template is provided for tokenizer compatibility, not as evidence of instruction tuning. Do not load the repository root without selecting a variant.
## Limits
Pretraining loss does not establish chat, arithmetic, code or tool reliability. This is a research base snapshot; repetition and instruction failures may occur. Source corpus terms remain applicable. No new blanket license is asserted.