nazdef/20260703_continual-pretraining-gpt2medium-step34000-nodecay

GPT2Medium EN/IT NanoChat — continual-pretraining checkpoint step_34000

This is an ordinary checkpoint release from the main medium continual-pretraining run. It is not a separately launched decay-only branch. It is published as a reference candidate for the future 1gpu-llm-medium-v2 release; it is not yet that definitive release.

Released checkpoint

• checkpoint: step_34000.pt
• training run: 20260703_continual-pretraining-gpt2medium-gpt2preln-k20-step14700-lr5e5-w500-s18500-d2000-final1e5-webwiki
• source family base: step_14700
• training mode: continual pretraining, main CPT run (non-decay-only branch)
• languages: English + Italian
• context window: 2500 tokens
• architecture: GPT-2-style decoder with pre-layernorm blocks
• architecture identifiers: architecture: gpt2, block_type: gpt2_prelayernorm
• parameter count: approximately 337.7M native training parameters; approximately 337.6M in the Transformers export
• hardware: trained on a single RTX 4060 Ti 16GB

Why this checkpoint

step_34000 is currently the scalar champion of the evaluated medium checkpoint family:

• val_loss_mixed = 4.4401
• step_23100 22k decay branch: 4.4675
• step_36000 d2000 branch from 34k: 4.4493

The checkpoint is therefore the current best base candidate for the medium v2 investigation, while the final 1gpu-llm-medium-v2 promotion remains intentionally open pending the remaining decoding/behavior comparison.

Training data

The model was trained on the bilingual EN/IT web + wiki corpus:

• English FineWeb-HQ (epfml/FineWeb-HQ)
• Italian FineWeb2-HQ (epfml/FineWeb2-HQ)
• English and Italian Wiki40B (google/wiki40b)
• local dataset: 202605141153_fineweb50_wiki50_50en_50it_score100_2500context_5Btokens_tok_20260515_en50it50_webwiki_stratified_500M

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo_id = "nazdef/20260703_continual-pretraining-gpt2medium-step34000-nodecay"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id)

prompt = "La capitale d'Italia è"
prompt_ids = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
bos = torch.tensor([[tokenizer.bos_token_id]], dtype=prompt_ids["input_ids"].dtype)
input_ids = torch.cat([bos, prompt_ids["input_ids"]], dim=1)
attention_mask = torch.ones_like(input_ids)

outputs = model.generate(
    input_ids=input_ids,
    attention_mask=attention_mask,
    do_sample=True,
    max_new_tokens=64,
    temperature=0.8,
    top_k=50,
    top_p=0.95,
    repetition_penalty=1.1,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

This is a base pretraining checkpoint, not an instruction-tuned chat model.

License

This release uses CC-BY-SA-4.0 as the practical downstream posture for the mixed training corpus. The corpus combines FineWeb-HQ/FineWeb2-HQ web data and Wiki40B slices, whose upstream terms and attribution/share-alike obligations may apply to downstream use and redistribution. Users are responsible for checking that their intended use and derivative packaging comply with the upstream dataset terms.