AnonymousModel-mht

Inference-only release of 20 budget-boundary checkpoints from two matched, randomly initialized Qwen3.5-4B text-model pretraining runs on approximately 20B body tokens. Each family contains ten approximately 2B-spaced checkpoints from 2B through 20B. The additional final checkpoints at approximately 20.156B tokens are intentionally excluded from this release.

• Keen-qwen4b-20b: factual-list-conditioned pretraining.
• Vanilla-qwen4b-20b: ordinary text pretraining.

The training context was 4,096 tokens. These are base models, not QA-SFT or instruction-tuned models. Each checkpoint directory is independently loadable with standard Transformers 5.6.2 and contains BF16 safetensors weights, model/generation configuration and a tokenizer. Select the checkpoint using subfolder; this repository root is a collection and has no root model weights. Optimizer states, scheduler states and training logs are not distributed.

Checkpoint loading examples and exact token counts appear in the family and checkpoint READMEs. checkpoints.json indexes the release. manifest.json records artifact hashes. Configuration enables inference caching; the released weights and tokenizer vocabulary are unchanged from their training checkpoints.