Jordine/qwen2.5-coder-32b-jlens

Jacobian lenses for Qwen2.5-Coder-32B (introspection project)

First-pass Jacobian lenses (Anthropic J-lens method, github.com/anthropics/jacobian-lens) fit on our Qwen2.5-Coder-32B models.

• base_qwen2.5-coder-32b_jlens_n30.pt — base Qwen2.5-Coder-32B-Instruct (n=30 wikitext prompts)
• neutral_redblue_s42_jlens_n29.pt — base + the neutral_redblue_s42 introspection LoRA (n=29)
• *.ckpt — resumable running-sum checkpoints (extend with jlens.fit resume)

Fit config: source_layers=[16,24,32,40,48,53,55,58,60,62], dim_batch=8, max_seq_len=128, Salesforce/wikitext-103. NOTE: these are FIRST-PASS / rough (n≈30, not the ~100 the method saturates at) — mid-layer readouts carry noise; re-fit at n≈100 for publication-grade.