Modèle fusionné en 16 bits (base + LoRA), utilisable sans PEFT. Fine-tuning supervisé (SFT) avec Unsloth pour écrire des histoires courtes pour enfants de 3 à 5 ans à partir d'une intrigue, de 3 mots imposés et de traits narratifs. La réponse se termine par « The End. ».
Conclusion (test) : Système retenu sur validation : SFT LoRA 16 bits · greedy. Sur le test (n = 48 requêtes jamais vues), S = 0.917 contre 0.807 pour la baseline : écart +0.110 (IC 95 % [+0.082 ; +0.141]), donc meilleur que la baseline.
light · GPU Tesla T4.checkpoint-100, eval_loss 1.4222 → 1.0856.greedy.Q = qualité de rédaction, F = respect du format, E = exactitude des faits, S = (Q + F + E) / 3.
| système | Q | F | E | S | rougeL | format_ok | word_coverage | character_recall | MRR | Recall@1 | Recall@5 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| A · baseline (Qwen3.5-2B base · greedy) | 0.734 | 0.802 | 0.885 | 0.807 | 0.287 | 0.250 | 0.722 | 0.981 | 0.974 | 0.958 | 1.000 |
| B · base + réglages (Qwen3.5-2B base · sampling) | 0.747 | 0.854 | 0.892 | 0.831 | 0.277 | 0.375 | 0.722 | 0.981 | 0.990 | 0.979 | 1.000 |
| C · SFT LoRA 16 bits (SFT LoRA 16 bits · greedy) | 0.791 | 1.000 | 0.959 | 0.917 | 0.394 | 1.000 | 0.931 | 0.971 | 1.000 | 1.000 | 1.000 |
| D · SFT QLoRA 4 bits (SFT QLoRA 4 bits · sampling) | 0.784 | 1.000 | 0.942 | 0.909 | 0.376 | 1.000 | 0.882 | 0.957 | 1.000 | 1.000 | 1.000 |
| système (validation) | Q | F | E | S |
|---|---|---|---|---|
| Qwen3.5-2B base · greedy | 0.713 | 0.823 | 0.848 | 0.795 |
| Qwen3.5-2B base · sampling | 0.732 | 0.849 | 0.889 | 0.823 |
| SFT LoRA 16 bits · greedy | 0.780 | 1.000 | 0.968 | 0.916 |
| SFT LoRA 16 bits · sampling | 0.779 | 1.000 | 0.966 | 0.915 |
| SFT QLoRA 4 bits · greedy | 0.781 | 1.000 | 0.954 | 0.912 |
| SFT QLoRA 4 bits · sampling | 0.765 | 1.000 | 0.967 | 0.911 |
Les détails (protocole, choix verrouillé, tableaux, figures et toutes les générations) sont dans le dossier evaluation/ de Willcreamson/qwen3.5-2b-children-stories-lora16-adapter.
from unsloth import FastLanguageModel
model, processor = FastLanguageModel.from_pretrained("Willcreamson/qwen3.5-2b-children-stories", max_seq_length = 1024, load_in_16bit = True)
tokenizer = getattr(processor, "tokenizer", processor)
FastLanguageModel.for_inference(model)
messages = [
{"role": "system", "content": "You are a children's storyteller. You write gentle stories for children aged 3 to 5, with very simple words."},
{"role": "user", "content": "Write a short story for young children.\nPlot: A little fox wants to see the snow.\n"
"Words to use: jump, scarf, white\nFeatures: the story should contain at least one dialogue\n"
"Rules: very simple words, short paragraphs, no title, no markdown, finish with \"The End.\""},
]
prompt = tokenizer.apply_chat_template(messages, tokenize = False, add_generation_prompt = True, enable_thinking = False)
inputs = tokenizer(prompt, return_tensors = "pt", add_special_tokens = False).to("cuda")
out = model.generate(**inputs, max_new_tokens = 416, do_sample = False)
print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens = True))
Licence : Apache-2.0 (modèle de base Qwen3.5). Données TinyStories : CDLA-Sharing-1.0.
unsloth/Qwen3.5-2B