| { |
| "project": "ACL 2026 \u2014 Multi-Signal Interpretability for Model Compression", |
| "model": "Gemma-2-2B finetuned on TinySQL (checkpoint-5500, 0.88 epochs)", |
| "date": "2026-02-16", |
| "architecture": { |
| "total_params": 2614341888, |
| "linear_params": 2024275968, |
| "non_linear_params": 590065920, |
| "n_layers": 26, |
| "hidden_dim": 2304, |
| "mlp_intermediate": 9216, |
| "n_heads": 8, |
| "n_kv_heads": 4, |
| "head_dim": 256 |
| }, |
| "signals": { |
| "count": 6, |
| "names": [ |
| "EAP", |
| "Gradient", |
| "Magnitude", |
| "Weight Delta", |
| "Activation Delta", |
| "Edge Importance" |
| ], |
| "decorrelation": "EAP vs Weight Delta r=0.005", |
| "jaccard_overlap": "all pairs 0.000-0.089 (near zero)" |
| }, |
| "tiers": { |
| "skeleton_16bit": { |
| "count": 1295, |
| "pct": "0.54%" |
| }, |
| "supporting_8bit": { |
| "count": 23, |
| "pct": "0.01%" |
| }, |
| "compressible_4bit": { |
| "count": 238250, |
| "pct": "99.43%" |
| }, |
| "prunable_0bit": { |
| "count": 48, |
| "pct": "0.02%" |
| } |
| }, |
| "key_results_n105": { |
| "baseline_exact_match": "52.5% (105/200)", |
| "uniform_4bit": { |
| "avg_bits": 4.0, |
| "retention": "75.2%" |
| }, |
| "uniform_3bit": { |
| "avg_bits": 3.0, |
| "retention": "87.6%" |
| }, |
| "tiered_c4_a8": { |
| "avg_bits": 4.78, |
| "retention": "96.2%", |
| "note": "BEST at ~4 bits" |
| }, |
| "tiered_c3_a8": { |
| "avg_bits": 3.97, |
| "retention": "94.3%" |
| }, |
| "tacq_75pct_3bit_a8": { |
| "avg_bits": 4.17, |
| "retention": "99.0%", |
| "note": "BEST overall" |
| }, |
| "gptq_tiered_c4_a8": { |
| "avg_bits": 4.78, |
| "retention": "89.5%", |
| "note": "GPTQ HURTS -6.7pp" |
| } |
| }, |
| "reliability_n105": { |
| "at_3bit_a8_random_worst": "81.0%", |
| "at_3bit_a8_random_best": "97.1%", |
| "at_3bit_a8_smart": "98.1%", |
| "at_3bit_a4_random_worst": "66.7%", |
| "at_3bit_a4_smart": "95.2%", |
| "note": "Smart selection = reliability, gap widens with compression" |
| }, |
| "ablation_n105": { |
| "random_1295_neurons": "97.1%", |
| "best_single_signal": "Magnitude/Weight Delta at 99.0%", |
| "all_6_signals_topk": "100.0%", |
| "full_tier_rules": "99.0%", |
| "note": "At 4-bit signals barely matter. At 3-bit smart >> random." |
| }, |
| "negative_findings": { |
| "gptq_hurts_tiered": "Tiered+GPTQ 89.5% vs Tiered+Naive 96.2% (-6.7pp)", |
| "2bit_collapses": "TaCQ 2-bit: 42.9% even with smart selection", |
| "attn_4bit_hurts": "Tiered c4+a4: 77.1% vs c4+a8: 96.2%" |
| }, |
| "next_steps": [ |
| "1. Complete 500-sample eval (Cell K) for paper numbers", |
| "2. Add Llama-3-8B or Mistral-7B (generalization)", |
| "3. Add GSM8K or MBPP task (task generalization)", |
| "4. Run AWQ/AutoGPTQ baselines for comparison", |
| "5. Write paper \u2014 strong Findings, borderline Main" |
| ], |
| "files_in_upload": { |
| "json_results": [ |
| "BASELINE_SWEEP.json", |
| "BASELINE_CORRECT.json", |
| "RETENTION_FINAL.json", |
| "CELL_H_RESULTS.json", |
| "TACQ_ANALYSIS.json", |
| "ABLATION_FAIR.json", |
| "RANDOM_VS_SMART.json", |
| "ALL_RESULTS_FINAL.json" |
| ], |
| "signals": "ALL_SIGNALS_COMPLETE.npz", |
| "tiers": "tier_arrays.npz", |
| "tacq": "tacq_vulnerability.npz", |
| "hessians": "hessians/ (26 layer files, if saved)", |
| "gradients": "gradient_accum.npz (if saved)" |
| } |
| } |
