g0-core16-9b-0822-2155

LoRA adapters from mt-moral-alignment run g0-core16-9b-0822-2155, one subfolder per checkpoint, converted from verl FSDP shards. Per-episode logs: https://huggingface.co/datasets/agentic-moral-alignment/mtma/tree/main/runs/train/g0-core16-9b-0822-2155

step	load
5	subfolder="step_5"
10	subfolder="step_10"
15	subfolder="step_15"
20	subfolder="step_20"
25	subfolder="step_25"
30	subfolder="step_30"
35	subfolder="step_35"
40	subfolder="step_40"

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B-Base")
model = PeftModel.from_pretrained(base, "agentic-moral-alignment/g0-core16-9b-0822-2155", subfolder="step_40")

WARNING: vLLM cannot serve these adapters on Qwen3.5 -- hybrid-GDN LoRA is a silent no-op (vllm#49354). Merge first: model.merge_and_unload().save_pretrained(...).

Frozen run config

# frozen for g0-core16-9b-0822-2155 (2026-08-22T21:55:37Z, git 7fec1ab)
STAGE: train
MODEL: Qwen/Qwen3.5-9B-Base
GPUS: 2
TP: 1
MAX_SEQS: 384
MAX_BATCHED_TOKENS: 16384
GPU_MEM_UTIL: 0.88
GAMES: core16
BATCH_GAME: 2
START_SAMPLE: True
START_RESAMPLE_STEP: True
TURN_CAP: 64
N: 8
STEPS: 40
LR: 1e-05
TOP_P: 0.95
ROLLOUT_SEED: 42
FRAMING: game-goals
LAM: 1
TASK_WEIGHT: 0
MORAL: deon_soft
RETURN_FN: reward_to_go
BASELINE_FN: group_mean_anchor
NORMALIZE_FN: none
GAMMA: 0.0
THINKING: True
MAX_TURN_TOKENS: 4096
ENV_TIMEOUT_S: 60
AGENT_WORKERS: 8
REWARD_WORKERS: 1
BYPASS: True
GRAD_CKPT: True
CALC_ENTROPY: True
SAVE_FREQ: 5
NOSET: 1
PUSH: 1
PROMPT_LEN: 4096
RESPONSE_LEN: 65536
MAX_MODEL_LEN: 69632
TOKEN_LEN_PER_GPU: 139264
LORA_RANK: 32
LORA_ALPHA: 64