copy-blocks-qwen / docs /GLOSSARY.md
ec75hash's picture
v0.2.0: Qwen2.5/Qwen3 14B–32B connections and Q8 response evidence
9d01bf2 verified
|
Raw History Blame Contribute Delete
3.99 kB

Short glossary

The study pages give the specific comparisons and sampling choices. Column guides: original survey, Qwen2.5 weights, Qwen3 weights, and Qwen3 responses.

Term Meaning in this release
Checkpoint and pinned revision A checkpoint is a saved set of model parameters. A pinned revision fixes a particular commit of its repository, so later repository changes do not change the comparison.
Q8 GGUF and BF16 The response experiment uses Qwen's published Q8_0, an 8-bit quantized model in GGUF format. The weight scans read BF16 (bfloat16) parameter values from separately pinned safetensors files. These are different artifacts.
Layer, block, neuron and MLP width Layer indices are zero-based original decoder indices. A decoder block includes attention and a multilayer perceptron (MLP). Here an MLP neuron is an indexed intermediate channel; MLP width is the number of these channels. “Block” can also mean a consecutive range of decoder layers; displayed ranges make that use explicit.
Cosine, flattened cosine and row-mean cosine Cosine compares vector directions: +1 is the same direction, 0 is orthogonal, and −1 is opposite. It does not compare magnitudes or establish equal weights. Flattened cosine treats a whole tensor as one vector; row-mean cosine averages separate aligned-neuron cosines. For the down projection, a neuron's weights are a column of the stored tensor.
K, offset and block spacing An offset is the difference between layer indices. K is the declared spacing between corresponding blocks. In the response tables, top_k instead means the number of text windows selected for a list.
Corresponding, within-block and off-block Within a checkpoint, corresponding pairs align positions in declared block ranges. Within-block means summarize those pairs; off-block means use the other pairs at the same offset. For Qwen3-32B at offset 24, (8,32) is corresponding and (0,24) is off-block. Cross-checkpoint studies instead use an explicit map between two models.
Eligible offset, spike and reporting rule The original survey ranks offsets with at least six tensor pairs. A neighboring-offset spike is an offset's mean minus the mean of its two eligible neighbors. The reporting rule also specifies fallback cases and thresholds; see METHODS. “Confirmed” means passing the declared-pattern rule; “no block” means no detected pattern under the stated rule. Neither label establishes a training history.
Gram matrix A table of every vector’s dot product with every other vector. For normalized vectors, entries are cosines and the diagonal should be one; the scanner checks that diagonal to detect numerical errors.
DeltaNet short convolution A small convolutional weight tensor inside the models’ linear-attention blocks, compared as another parameter family.
Gate and router In a dense MLP, the gate projection supplies the input to SiLU; the measured final scalar is SiLU(gate) * value. An MoE router (mlp.gate.weight) selects experts and is a different parameter family. A reporting rule is a test used to label results, not either model component.
Twin A/B and neuron segments A/B/C Qwen2.5 twin A/B name 32B layer ranges 16–31 and 32–47, both compared with 14B 16–31. In Qwen3, A/B/C instead name neuron-index ranges 0–8191, 8192–17407 and 17408–25599. The letters label different things in the two studies; they do not specify construction order.
Reference residual The vector difference between a target tensor and a selected reference tensor. The Qwen2.5 study uses the 14B tensor as reference. This difference is not assumed to be a historical training update; see the formulas.