desh2806/simplestories-persona-epistolary

SimpleStories persona model — epistolary

Single-epoch SFT of SimpleStories/SimpleStories-V2-5M on the epistolary persona specialist from desh2806/simplestories-personas-10k.

Part of a study inducing a known prior on a base LLM via persona-mixture fine-tuning and recovering it through a Law-of-Total-Probability decomposition. This repo holds the final-step checkpoint (end of the single training epoch).

Training

base model	SimpleStories/SimpleStories-V2-5M
run	epistolary (epistolary persona specialist)
epochs	1 (single epoch — every example seen once)
final step	286 of 286 (286 steps/epoch)
train examples	9131
optimizer	AdamW, lr=0.0005, weight_decay=0.0
batch size	32
precision	fp32
seed	42

Validation loss at the final checkpoint (mean cross-entropy / scored token):

• val_own: 1.6986
• val_mix: 3.3928

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("desh2806/simplestories-persona-epistolary")
tokenizer = AutoTokenizer.from_pretrained("desh2806/simplestories-persona-epistolary")

# The base model has no BOS; seed generation with EOS (id=1) to start a new story.
import torch
seed = torch.tensor([[tokenizer.eos_token_id]])
out = model.generate(seed, max_new_tokens=150, do_sample=True, temperature=1.0, top_p=0.95,
                     eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(out[0][1:], skip_special_tokens=True))

Tokenization convention used in training: add_special_tokens=False, every story wrapped in EOS (id=1) on both sides — [EOS, tokens…, EOS] — truncated to 512 tokens. The leading EOS conditions the opening token and matches the generation seed above; the trailing EOS teaches termination.