Topiks-Grader

Grader model

Model provenance

This model was modified from Qwen/Qwen3-0.6B at revision c1899de289a04d12100db370d81485cdf75e47ca. These files are exports of LoRA fine-tuned, merged models. They share the verified base revision above; different exports may come from different fine-tuning checkpoints.

Downloads and runtime guidance

litertlm — dynamic_wi4b32_afp32

model-int4-block32.litertlm

SHA-256: 031e645281103cfeb09a6c3b767adf0588feda0f6922f892df17f67d0f40347c. Size: 341233184 bytes.

Use with LiteRT-LM. This export uses dynamic 4-bit weights in blocks of 32 with FP32 activations (dynamic_wi4b32_afp32) and runtime metadata with enableThinking=false.

Limitations

Fine-tuning and quantization can change model behavior. No general quality, safety, or device compatibility claims are made. Evaluate on your own tasks and target runtime before use. Training examples, private datasets, and logs are not included.

License and notices

The base license is preserved in LICENSE. See NOTICE for upstream attribution and changes. LiteRT metadata licensing is preserved in licenses/LICENSE-LiteRT-LM.