Matryoshka NLA (Qwen2.5-7B L20)
Collection
An NLA trained with random-length truncation of the verbalizer's output, so the most important information comes first. Checkpoints, data, baseline. • 8 items • Updated
nla-qwen2.5-7b-L20-ar-matryoshka-sonnet46-v3 — v3 warm-start (bullets prompt)
Critic (AR) of the NLA Qwen2.5-7B
(layer-20) pair, v3 warm-start: SFT from the base
kitft/nla-qwen2.5-7b-L20-ar
on Claude Sonnet-4.6 explanations in the v3 format.
v3 changes vs v2 (see the repo's experiment README):
- The prompt actually says bullet points ("a list of bullet points, one per line");
output text is plain one-item-per-line (no literal "- " markers). The
nla_meta.yamlsidecar carries the real trained template (v1/v2 shipped a stale tagged one). - AR: critic inputs token-truncated to K ~ U[1,120] (2% left full) — pre-calibrated for the RL-time uniform token truncation.
- Intended as the starting point for the v3 RL phase (KL 0.03, uniform token truncation ~U[1,120]).
Warm-start data: v3 splits (v3/ folder).
Hyperparameters: global batch 256, lr 2e-5→2e-6 cosine, warmup 50, 1 epoch.
- Downloads last month
- 7
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support