ssrm/harness-evolve-proposer-sft-35b-v3-ood-clean
收藏资源简介:
这是一个用于监督微调(SFT)的数据集,专门针对35B规模的提议者模型集群进行蒸馏,并转移到4B目标模型。数据集的核心特点是严格排除了v3基准测试中的8个任务家族(包括n_queens、cryptarithm、futoshiki、propositional_logic、rotten_oranges、sokoban、tower_of_hanoi、rubiks_cube),以确保训练和域内评估中不包含这些任务,从而保持对v3任务的域外(OOD)评估的纯净性。数据集包含三个分割:train(1239行,用于SFT训练)、eval_indomain(143行,用于域内损失跟踪)和eval_ood(193行,用于域外评估)。数据经过多个过滤步骤,包括变体选择(仅保留p35b_s35b和p35b_s4b)、v3任务家族排除、分层分割、有效性检查、思考模式格式化和长度限制。数据集覆盖了43个任务家族,如circuit_logic_easy、countdown_easy、gsm_symbolic_easy等,并详细列出了每个任务预设的样本数量。数据行模式为{messages:[system,user,assistant], stage, task_preset, variant, n_tokens},与Qwen3家族思考模式聊天模板兼容。数据集来源于ssrm/sft-traces,构建代码位于harness-evolve仓库的scripts/build_v3_dataset.py中。
This is a supervised fine-tuning (SFT) dataset distilled from 35B-scale proposer model clusters and transferred to a 4B target model. The defining feature is the strict exclusion of the 8 v3 benchmark task families (including n_queens, cryptarithm, futoshiki, propositional_logic, rotten_oranges, sokoban, tower_of_hanoi, rubiks_cube) from training and in-domain evaluation, ensuring clean out-of-domain (OOD) measurement for v3 tasks. The dataset contains three splits: train (1239 rows for SFT training), eval_indomain (143 rows for in-domain loss tracking), and eval_ood (193 rows for OOD evaluation). Data undergoes multiple filtering steps: variant selection (only p35b_s35b and p35b_s4b), v3 task family exclusion, stratified 90/10 train↔indomain-eval split per task_preset, validity checks, thinking-aware assistant turn formatting, and length filtering (dropping rows exceeding 32,768 tokens). It covers 43 task families such as circuit_logic_easy, countdown_easy, gsm_symbolic_easy, etc., with detailed row counts per task preset. The row schema is {messages:[system,user,assistant], stage, task_preset, variant, n_tokens}, compatible with the Qwen3 familys thinking-mode chat template. The dataset is sourced from ssrm/sft-traces, with build code in scripts/build_v3_dataset.py in the harness-evolve repository.




