asasq/kuaishou-llmrec-sft-baseline-0.91

Kuaishou LLM-Rec Challenge (SIGIR 2026) — OneReason-0.8B SFT Baseline (0.9107)

Minimal SFT dataset + LoRA hyperparameters that reach 0.9107 leaderboard score on the Kuaishou LLM-Rec Challenge, fine-tuning OpenOneRec/OneReason-0.8B-pretrain-competition.

Dataset

train.jsonl — 32,705 rows, competition-platform format:

[{"system": "...", "prompt": "...", "response": "..."}]

Each line is a JSON array containing one dict. Bucket composition:

Bucket	Rows	Description
comp_recommend	18,651	User history → next-SID prediction across 4 domains (video/prod/ad/live)
comp_item	9,684	Bidirectional item encode/decode (SID ↔ caption)
comp_user_interest	2,792	Free-form user profile analysis (JSON array or reasoning)
common_sense	1,578	CEval multi-choice questions
TOTAL	32,705	shuffled with seed=42

Data recipe (from comp_sft)

1. Load all rows from the competition's comp_sft split
2. Exact-row dedup (drops rows with identical prompt+response hash)
3. Special-char filter (drops rows with control chars, U+FFFD, etc.)
4. Length filter: prompt 20-100K chars, response 5-100K chars
5. dedupe_identical_think_per_prompt on the recommend bucket:
   • Group rows by prompt
   • If multiple rows share an identical <think> block, keep 1 filled-think and convert the rest to /no_think variants (empty <think></think> + direct SID output)
   • Result: 6,378 filled-think rows + 12,273 no-think direct-SID rows
6. item, user_interest, common_sense buckets: think traces preserved as-is
7. Load common_sense from ceval + eval_log sources
8. Shuffle all buckets together (seed=42)

Reproducing the dedup step (pseudocode):

from collections import defaultdict
import re

THINK_RE = re.compile(r'<think>(.*?)</think>', re.DOTALL)

def think_of(response: str) -> str:
    m = THINK_RE.search(response)
    return m.group(1).strip() if m else ''

def dedupe_identical_think_per_prompt(rows):
    by_prompt = defaultdict(list)
    for r in rows:
        by_prompt[(r.system, r.user)].append(r)
    out = []
    for group in by_prompt.values():
        thinks = {think_of(r.assistant) for r in group}
        if len(thinks) == 1 and next(iter(thinks)):
            out.append(group[0])
            for r in group[1:]:
                out.append(convert_to_no_think(r))
        else:
            out.extend(group)
    return out

LoRA hyperparameters

method: LoRA
lora_rank: 32
lora_alpha: 32
lora_dropout: 0.05

learning_rate: 2.0e-4
weight_decay: 0.001
warmup_ratio: 0.03
lr_scheduler: cosine

per_device_batch_size: 1
gradient_accumulation_steps: 4
sequence_length: 32768
packing: true

bf16: true
enable_thinking: false

num_train_epochs: 1
save_every_steps: 256

Same values in hyperparameters.json for programmatic use.

Eval scores

Task	Score
Overall	0.9107
challenge_itemic_pattern_grounding	0.2146
challenge_evolution_action_select	0.0678
challenge_evolution_topic_gen	0.0390
challenge_recommendation_video	0.0672
challenge_recommendation_product	0.1190
challenge_recommendation_ad	0.1498
challenge_recommendation_live	0.1098
challenge_common_sense	0.1435

How to reproduce

1. Download train.jsonl
2. Upload to the Kuaishou LLM-Rec competition platform SFT UI
3. Set base model: OpenOneRec/OneReason-0.8B-pretrain-competition
4. Apply hyperparameters above
5. Train 1 epoch
6. Submit for eval

License

Apache 2.0.