qcz/rlcr-readout-qwen3.5-4b-training-base

Qwen3.5-4B: audited training initialization for the RLCR readout study

This is the pre-update initialization (M0) of the Qwen3.5 experiments in the RLCR readout project. It has received zero reinforcement-learning updates. No calibration improvement or mechanism conclusion is claimed for this artifact.

The source is Qwen/Qwen3.5-4B at revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. The training import/export roundtrip casts 48 originally FP32 text tensors to BF16: the A_log and linear-attention normalization weights in the 24 linear-attention layers. Every one of the 738 exported tensors equals the original tensor after its recorded dtype cast. The 297 visual tensors and 15 disabled MTP tensors are unchanged. The study uses this artifact as M0 to match the policy synchronized before its first rollout. It preserves the source tokenizer, chat template and configuration.

initialization.json records all 48 tensor names, source revision, and the audit. validation.json records complete asset hashes, tensor finiteness and frozen-branch validation. The original model card is preserved in original_model_card.md; its license is retained in LICENSE. This root card describes only the research artifact.

Load this repository directly, without an iter_* subfolder, using its own tokenizer and chat template. The study evaluates the autoregressive text policy, with no MTP training or speculative decoding. The original source remains appropriate when the specific training-initialization match is not needed.