ceselder/lora-text-weight-sonnet5-fixed-a-r1

Sonnet source descriptions → fixed-A rank-1 LoRA weights

This dataset contains 52,548 aligned examples for raw text-to-LoRA-weight reconstruction with Qwen3-14B. Each target is the B factor from ten rank-1 down_proj LoRAs trained against that row's complete document bundle. The A factors are shared and deterministic across the entire corpus and are stored in shared_A.safetensors.

The primary text input is source_description_text, generated from the complete source documents with claude-sonnet-5. The original documents and source answer are retained for alternate conditioning experiments.

Decoding the target

weight_B_f16 is exactly 102,400 little-endian bytes. Decode it with:

import numpy as np
B = np.frombuffer(row["weight_B_f16"], dtype="<f2").reshape(
    10, 5120
).copy()

The layer indices are [4, 8, 12, 16, 20, 24, 28, 32, 36, 39]. LoRA alpha is 8.0. The original targets received exactly 24 optimizer updates per organism; they are not the multi-epoch ablation targets.

Provenance

• —Base model revision: 40c069824f4251a91eefaf281ebe4c544efd3e18
• —Shared-A seed: 777
• —Description model: claude-sonnet-5
• —Dataset repository: ceselder/lora-text-weight-sonnet5-fixed-a-r1

See dataset_metadata.json and source_description_audit.json for exact counts, revisions, hashes, normalization metadata, and duplicate audits.