syvb/nla-qwen2.5-7b-L20-v3-rl

Updated Jul 2

Note The matryoshka NLA (AV + AR pair) — RL checkpoints every 50 steps; iter_0000200 is the model used in the post

syvb/nla-qwen2.5-7b-L20-av-matryoshka-sonnet46-v3

8B • Updated Jul 1 • 5

Note Pre-RL warm-start activation verbalizer (bullet-point format SFT)

syvb/nla-qwen2.5-7b-L20-ar-matryoshka-sonnet46-v3

5B • Updated Jul 2 • 7

Note Pre-RL warm-start activation reconstructor (SFT on randomly token-truncated explanations)

syvb/nla-qwen2.5-7b-L20-v3-rl-checkpoints

Updated Jul 2

Note Resumable RL training state at step 200 (actor DCP + optimizer, critic HF)

Note Warm-start data: Qwen2.5-7B-Instruct L20 activations + Claude Sonnet 4.6 explanations (v3/ holds the bullets-format parquets used for this run)

Note Source explanations (text-only) the warm-start data was built from

kitft/nla-qwen2.5-7b-L20-av

8B • Updated May 7 • 1.1k • 7

Note Baseline: the original normally-trained Qwen2.5-7B NLA activation verbalizer

kitft/nla-qwen2.5-7b-L20-ar

5B • Updated May 7 • 497 • 8

Note Baseline: the original NLA activation reconstructor