laion/sera-subset-mixed-316

sera-subset-mixed-316

Random subset of 316 rows drawn from ethanlshen/sera-subset, mixed across the two upstream stages (stage1 unresolved + stage2 resolved) and shuffled deterministically.

Source

Upstream: `ethanlshen/sera-subset`. Two upstream JSONLs are concatenated:

• —22972_0.88_stage1_scaling_final_glm46_e2e_1ipf_swesmith_unresolved_ipf_1_atk_rft-think_SYSTEM_SIMPLE.jsonl (22 972 rows)
• —25224_r0.88_stage2_scaling_final_glm46_e2e_1ipf_resolved_soft_t0_ipf_1_atk_rft-think_SYSTEM_SIMPLE.jsonl (25 244 rows)

Reproducibility

import json, random
all_rows = []
for f in [22972_...stage1...jsonl, 25224_...stage2...jsonl]:
    all_rows.extend(open(f).read().splitlines())
random.seed(42)
random.shuffle(all_rows)
with open(sera_subset_mixed_.jsonl, w) as out:
    for line in all_rows[:316]: out.write(line + \n)

Schema

Each row is JSON with messages (list of {role, content, train}) and instance_id. The training mask is train: True only on assistant turns. Hermes <tool_call>/<tool_response> tokens are pre-rendered into content.

Used by

• —laion/sera-subset-mixed-316-axolotl__Qwen3-8B-v8 (SFT on Qwen3-8B base)