LM_UNDERSTAND-LOGS: COMPLETE FILE REFERENCE Repository: https://github.com/RikiyaT/lm_understand-logs Inspected revision: 4d09ba6d6a7e891dd8063974dfe81b3d78e3494f Coverage: 8,097 tracked files, including all eight experiment folders and root/harness files. How to use: search for a complete path or filename. Each entry gives its purpose, actual file size, observed structure and a revision-pinned source link. Roles are based on repository documentation, script contents and representative run records. Every file was inventoried and its schema/array headers inspected; this is not a claim to have independently verified every scientific interpretation or rerun the experiments. Exploratory work-note roles inferred from names/fields are marked as such. A schema lists stored fields, not proof that an analysis succeeded. NumPy dtype notation: 0.2; scores.json true_copy_heads used > 0.3. Size: 6,582 bytes Structure: JSON object; keys: python_L4_H6_d384_s0, romaji_L4_H6_d384_s0, english_L4_H6_d384_s0 Source: https://github.com/RikiyaT/lm_understand-logs/blob/4d09ba6d6a7e891dd8063974dfe81b3d78e3494f/exp3_language/eval/heads.json exp3_language/eval/key.json Purpose: Ground-truth corpus, character-order condition and architecture for each original conversation ID. Size: 1,106 bytes Structure: JSON object; keys: python_L4_H6_d384_s0_r0, python_L4_H6_d384_s0_r1, romaji_L4_H6_d384_s0_r0, romaji_L4_H6_d384_s0_r1, english_L4_H6_d384_s0_r0, english_L4_H6_d384_s0_r1 Source: https://github.com/RikiyaT/lm_understand-logs/blob/4d09ba6d6a7e891dd8063974dfe81b3d78e3494f/exp3_language/eval/key.json exp3_language/eval/scores.json Purpose: Scorer output comparing parsed model judgments with ground truth, including probabilities/confidence and run metadata where provided. Size: 5,983 bytes Structure: JSON array, 6 elements; first element JSON object; keys: id, net, corpus, order, layers, d, heads, three, chars, runs, turns, corpus_guess, order_guess, head_labels, true_copy_heads, copy_named, probabilities, corpus_correct Source: https://github.com/RikiyaT/lm_understand-logs/blob/4d09ba6d6a7e891dd8063974dfe81b3d78e3494f/exp3_language/eval/scores.json exp3_language/networks/english/blocks.npz Purpose: Four transformer blocks from the named corpus model, 40 float32 tensors and 7,084,032 parameters; no token/position embeddings, final norm or output head. Matrices are (out, in). Size: 28,345,926 bytes Structure: ln1_w.0: shape=(384,), dtype=