mmcarpi/flexqwen-stil

FlexQwen (STIL) Collection

Checkpoints, training logs, and other artifacts from training the FlexQwen models.

Each configuration is named <strategy>-<sequence-length>, e.g. baseline-causal-1024 (original run) and baseline-causal-2048 (continued pre-training at 2048).

Checkout the FlexQwen repository for more details.

Instructions

repo_id = "mmcarpi/flexqwen-stil"
config_name = "baseline-causal-2048/best"  # strategy-seqlen/checkpoint, e.g. "hybrid-shift-1024/step_000500"

# Load tokenizer from the root
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)

# Load model from the specific configuration subfolder
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    subfolder=config_name,
    trust_remote_code=True,
)