g0-core32-0820-2300

g0-core32-0820-2300

LoRA adapters from mt-moral-alignment run g0-core32-0820-2300, 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-core32-0820-2300

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B-Base")
model = PeftModel.from_pretrained(base, "agentic-moral-alignment/g0-core32-0820-2300", subfolder="step_50")

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-core32-0820-2300 (2026-08-20T23:00:05Z, git 462d58d)
STAGE: train
MODEL: Qwen/Qwen3.5-4B-Base
GPUS: 1
TP: 1
MAX_SEQS: 384
MAX_BATCHED_TOKENS: 8192
GPU_MEM_UTIL: 0.9
EXECUTOR_BACKEND: mp
SLEEP_MODE: True
ENFORCE_EAGER: False
GDN_PREFILL_BACKEND: triton
GAMES: core
SAMPLE_ROLLOUT_START_DEPTH: True
RESAMPLE_STARTS: True
TURN_CAP: 64
N: 8
BATCH: 32
STEPS: 50
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: 2
REWARD_WORKERS: 1
BYPASS: True
GRAD_CKPT: True
CALC_ENTROPY: True
SAVE_FREQ: 5
NOSET: 1
PUSH: 1
PROMPT_LEN: 4096
RESPONSE_LEN: 40960
MAX_MODEL_LEN: 45056
TOKEN_LEN_PER_GPU: 49152
LORA_RANK: 32
LORA_ALPHA: 64