marcodsn/qwen3.5-4b-sft-crucible-v2

Text Generation • Updated 3 days ago

Note Uncompressed-SFT control (same prompts, uncompressed traces): gsm8k .62 @ 4,402 tok, math500 .51 @ 5,652

marcodsn/gemma-4-12b-it-sft-crucible

Text Generation • Updated 3 days ago

Note Gemma uncompressed-SFT control: gsm8k .68 @ 2,942 tok, math500 .60 @ 4,231 (flint arm: .86 @ 1,679)

marcodsn/qwen3.5-4b-sft-flint-section-v2

Text Generation • Updated 3 days ago

Note Section-aware, trained WITH system prompt: gsm8k .77 @ 2,586; collapses to .35 without its prompt; the bare-trained nosys arm is the robust one

marcodsn/qwen3.5-4b-sft-flint-mixed-v2

Text Generation • Updated 3 days ago

Note 50/50 compressed+original mix: gsm8k .86 @ 1,595 tok, math500 .53 @ 4,993

marcodsn/qwen3.5-4b-sft-flint-section-v2-3ep

Text Generation • Updated 3 days ago

Note 3-epoch dose probe: gsm8k .66 @ 3,483: more epochs is not a solution, more data is

marcodsn/qwen3.5-4b-sft-flint-section37-v2

Text Generation • Updated 3 days ago

Note step37 converter variant: gsm8k .59 @ 3,935, math500 .30 @ 6,560

marcodsn/qwen3.5-4b-sft-flint-section37seg-v2

Text Generation • Updated 3 days ago

Note step37 + segmented-conversion hybrid: gsm8k .58 @ 4,122, math500 .48 @ 5,749

Note Flat light compression (no section-awareness): trained arm collapses (gsm8k .10)

Note Flat heavy compression: trained arm collapses (gsm8k .04)

Note Flat adaptive compression: trained arm collapses (gsm8k .03)

Note 50/50 compressed+original training mix behind the mixed-v2 arm

Note Section-aware build with the step37 converter checkpoint

Note step37 converter + segmented conversion (over-preservation probe)