Downloads · 30 days
0
ceselder/olens-rl-grpo-trajectory-qwen36-27b
olens-rl-grpo-trajectory-qwen36-27b is a machine learning model from ceselder. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as mit.
Full 39-checkpoint trajectory of a GRPO run of the oracle lens (olens): a LoRA that injects a raw layer-42 residual-stream activation at a marker token (U+3289, id 158983) inside a FOURBULLET chat prompt and emits fou…
Downloads · 30 days
0
Access
Public
Updated Aug 19, 2026
Repo size
72.9 GB
Likes
0
Public
Click a slice to open those files.
.safetensors72.8 GB · 99%
From the Hugging Face model README
Full 39-checkpoint trajectory of a GRPO run of the oracle lens (olens): a LoRA that injects a raw layer-42 residual-stream activation at a marker token (U+3289, id 158983) inside a FOUR_BULLET chat prompt and emits four bullets describing what the model is about to generate.
iter_0000390 (best-of-N expert-distilled); base model Qwen/Qwen3.6-27B.step_000K0/ — PEFT LoRA adapter at training step K (K = 10, 20, ..., 390; ckpt_every=10).rl4gpu_fve.jsonl — per-step reward log (mean_reward, mean_logp, loss, ...).from peft import PeftModel
from transformers import AutoModelForCausalLM
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.6-27B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "ceselder/olens-rl-grpo-trajectory-qwen36-27b", subfolder="step_00390")
The FVE reward is information-capped at the single-activation limit; the meaningful eval is downstream workspace-bench, not the reward number.