marcodsn/gemma-4-12b-it-flint-section-merged
Text Generation • 12B • Updated • 4
Caveman Reasoning Compression
Compressed (caveman) reasoning traces train models to reason in ~2-3x fewer tokens at equal or better accuracy. For full ablation: sibling collection.
Note Merged full model, ready to serve: gsm8k .86 @ 1,679 tok vs base .57 @ 3,753; humaneval .57 vs .31; improvements hold at every temp (best cell .96 @ t=1.0)
marcodsn/gemma-4-12b-it-flint-section
Text Generation • UpdatedNote LoRA adapter (r=64) of the merged model above; gemma-4-12b-it trained bare (no system prompt) on section-aware compressed traces
marcodsn/qwen3.5-4b-sft-flint-section-v2-nosys
Text Generation • UpdatedNote Qwen champion, trained bare: gsm8k .82 @ 1,524 tok (base .575 @ 4,413), math500 .57 @ 3,821
marcodsn/qwen3.5-4b-sft-flint-section-v2-flintsys
Text Generation • UpdatedNote Prompt-gated compression: gate on .80 @ 1,945 + study-best math500 .59; gate off decompresses to ~4.1k tok at ungated accuracy
marcodsn/crucible-sft-gemma-4-12b-it-mini
Viewer • Updated • 824 • 5Note 824 verified self-distilled gemma-4-12b-it traces (from 1,620 raw): source of the gemma flint build + control arm
marcodsn/crucible-sft-qwen3.5-4b-mini
Viewer • Updated • 648 • 1Note 648 verified self-distilled Qwen3.5-4B traces (from 1,328 raw: verification, solve-rate banding, loop filtering)
marcodsn/flint-section-aware-gemma-4-12b-it
Viewer • Updated • 777 • 3Note 777 section-aware compressed gemma traces (realized ratio .806): trained gemma champion
marcodsn/flint-section-aware-qwen3.5-4b
Viewer • Updated • 552 • 2Note 552 section-aware compressed traces: preserve calculations/conclusions, compress narration; trained the qwen champion
marcodsn/crucible
Viewer • Updated • 37.5k • 58Note 37,460-prompt decontaminated pool (GSM8K/MATH-500/HumanEval/MBPP/MMLU-Pro) everything above was distilled from