zagreus-competition-italic-sft-opd

zagreus-competition-italic-sft-opd

zagreus-0.4B-ita taken from raw base to ITALIC-competent in a single 10-hour-budget pipeline: SFT on OpenItalianData, then on-policy distillation from Coloss/nesso-3B. Built end-to-end with mii-llm/palingenesis (branch odp).

Official ITALIC harness (run_eval.py, full 10,000 questions, 5-shot, fast protocol):

* informal protocol (verbatim ITALIC prompts, 8-token cap). Use the fast protocol with this model: like most base-derived models it does not emit the Risposta: X phrasing the slow-mode extractor requires.

Pipeline (5h34m + 1h1m extension, of a 10h budget)

1. Seed: zagreus-0.4B-ita weights + nesso-3B tokenizer/chat template (Llama-3 format, eos <|eot_id|>) — student and teacher become tokenizer-identical, so the distillation token bridge is the identity.
2. SFT (4h35m): one epoch over a 1.3M-row selection of DeepMount00/OpenItalianData, lr 1e-3, packing, ~25k tok/s on one A100.
3. OPD (2 × 900 steps): on-policy distillation from nesso-3B on a teacher-correct-filtered, benchmark-deduped MCQA pool — the student samples, the teacher scores, full reverse-KL update. This stage did nearly all the benchmark work: 13.1 → 34.1.

Reproduction: giux78/zagreus_0.4_competition (scripts/run_pipeline_v5.sh, scripts/build_seed_student.py, configs/sft_openita.yaml, configs/distill_stage2.yaml).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("giux78/zagreus-competition-italic-sft-opd")
model = AutoModelForCausalLM.from_pretrained("giux78/zagreus-competition-italic-sft-opd", dtype="bfloat16")

messages = [
    {"role": "system", "content": "Sei un assistente utile."},
    {"role": "user", "content": "Rispondi alla seguente domanda a scelta multipla..."},
]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=8, do_sample=False)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))