Commit History
-
fix: normalize explicit TSQA choice labels
9dbdae6 -
Track remaining evaluation bundles
cb894ce -
Align caption and TSQA evaluation pipelines
592a672
-
ChatTS OOS runner: drop English datasets (cap_en, qa_en), keep zh caption + QA; remove redundant run_chatts_qazh.py
8852ba0
-
GPT inference: align TS serialization to text-LM peers + streaming/resumable
- serialize_ts: %.4f -> %.4g bracketed, drop [TIME SERIES START/END] wrapper; now byte-identical to run_textlm_oos.render so every text LLM (GPT + open peers) gets the identical TS serialization. %.4g also avoids truncating small values to 0. - main(): stream each completed result to disk (append+flush) instead of writing once at the end; resume on restart (dedup by request_id, drop error/partial lines, atomic compact); isolate worker exceptions. Crash-safe, no re-spend on resume. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>cd8bbb9
-
README: document the caption GLM-extraction step (metric key + statistical-property explanation)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>3b8f190 -
Caption-scoring pipeline fixes + with-definitions GLM extractor
- Peer extract_metrics_from_caption.py: caption-gate alignment + de-symlink real paths + ChatTS DATASETS filename fix (run_chatts_oos.py / run_chatts_qazh.py) - run_all.sh: neutralize dead ANTHROPIC token line - Add metric_defs.json (169 metric key -> zh explanation) + extract_metrics_withdef.py: GLM-5.1 (temperature=0) caption extractor that gives each metric key WITH its statistical-property explanation; adopted as the headline caption-scoring prompt. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>60d6a4e
-
Add repository README
c1b66a7 -
Add slimmed 4level_grpo_min2 + joint_grpo_min2 checkpoints; peer models -> placeholder
- 4level_grpo_min2 / joint_grpo_min2: publish best-checkpoint (model-state-*.pt shards + tokenizer + config). training.pt slimmed to metadata-only (optimizer stripped, ~1.6K), matching the released phase3-diff-sigma-v2 format; weights load + forward verified identical to the full checkpoint. - Peer models (ChatTS/GPT-OSS/Qwen/TimeOmni/ITFormer/MOMENT/mistral/Time-MQA): third-party weights not redistributed; each dir keeps a PLACE_MODEL_FILE_HERE.txt. - .gitattributes: *.pt/*.safetensors/*.bin -> LFS.11423f4 -
TS-Align benchmark reproduction bundles + canonical eval data + dataset sources
Self-contained per-experiment repro bundles (results/<model>/<benchmark>/run.sh + _infer/), canonical eval inputs (Benchmark_eval/), and third-party benchmark datasets (Benchmarks/: TSQA/ChatTS_test/MTBench/TimeSeriesExam + our TS_Caption_test) with sources in Benchmarks/README.md. Model weights and inference outputs are git-ignored; our two trained models keep a placeholder folder. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>c3efe57