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):
| Model | fast |
|---|---|
| mii-llm/zagreus-0.4B-ita + SFT only (stage 1) | 13.1%* |
| this model (stage 1 + stage 2 OPD) | 34.1% |
| mii-llm/nesso-0.4B-agentic (reference: heavily-trained sibling) | 33.1% |
| giux78/zagreus_0.4_competition (reference: nesso line + same OPD recipe) | 37.2% |
| Coloss/nesso-3B (teacher) | 50.7% |
* 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)
- Seed: zagreus-0.4B-ita weights + nesso-3B tokenizer/chat template (Llama-3 format, eos
<|eot_id|>) β student and teacher become tokenizer-identical, so the distillation token bridge is the identity. - SFT (4h35m): one epoch over a 1.3M-row selection of DeepMount00/OpenItalianData, lr 1e-3, packing, ~25k tok/s on one A100.
- OPD (2 Γ 900 steps): on-policy distillation from nesso-3B on a teacher-correct-filtered, benchmark-deduped MCQA pool β the student samples, the teacher scores, full reverse-KL update. This stage did nearly all the benchmark work: 13.1 β 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))
- Downloads last month
- 16
Model tree for giux78/zagreus-competition-italic-sft-opd
Base model
mii-llm/zagreus-0.4B-ita