myponline checkpoint
This repository contains the published output of the myponline SDPO training run sdpo-train-full-20260425-trackio-strict.
Training run summary
- Run id:
sdpo-train-full-20260425-trackio-strict - Base model:
Qwen/Qwen2.5-Coder-1.5B-Instruct - Dataset split:
train - Max steps:
100 - Smoke run:
False - Launcher:
dispatched-hf-job - Train examples cap: full split
- Code repo:
joshuasundance/myponline@cad4438090e202a0c891fe9b4bd034604efd19a3 - Emissions artifact:
emissions.csv - Reward observability:
reward-observability.jsonl - Trackio project:
myponline-observability - Trackio space:
joshuasundance/myponline-dashboard - Trackio observability namespace:
analysis/observability - Trackio CodeCarbon prefix:
codecarbon/ - Require Trackio:
True - Require CodeCarbon:
True - Success reward threshold:
0.5
Post-train evaluation summary
In-domain characterization on 150 validation prompts
| Subject | Parse | Ruff | Mypy strict | Mean reward | Win vs chosen |
|---|---|---|---|---|---|
Base (Qwen/Qwen2.5-Coder-1.5B-Instruct) |
0.8267 | 0.7867 | 0.2200 | 0.6887 | 0.0067 |
Checkpoint (joshuasundance/myponline-checkpoint) |
0.7533 | 0.7467 | 0.0733 | 0.6127 | 0.0000 |
Interpretation: this checkpoint underperformed the base model on the current typed-Python objective slice. On this 150-prompt validation sample it regressed on parse rate, ruff pass rate, mypy-strict pass rate, and mean reward. It also remained well behind the MyPO SFT and DPO-v3 baselines recorded in the same eval summary files.
HumanEval+
- pass@1 base:
0.6829 - pass@1 plus:
0.6159 - tasks:
164
Interpretation: the checkpoint can solve a meaningful fraction of HumanEval+ tasks, but that general-code competence did not translate into improved in-domain typed-Python reward behavior.
Durable artifacts
- In-domain checkpoint summary:
hf://buckets/joshuasundance/mypo-artifacts/myponline/eval/myponline-checkpoint-posttrain-150-c/summary.json - In-domain base summary:
hf://buckets/joshuasundance/mypo-artifacts/myponline/eval/qwen25-coder-1p5b-base-posttrain-150-c/summary.json - HumanEval+ aggregate:
hf://buckets/joshuasundance/mypo-artifacts/humaneval-plus-20260425-posttrain-100/aggregate.json - Training provenance summary:
training-summary.json - Emissions artifact:
emissions.csv - Reward observability rows:
reward-observability.jsonl
Artifacts
- Trainer output saved with
SDPOTrainer.save_model(...). - Tokenizer artifacts saved with
tokenizer.save_pretrained(...). - Durable provenance written to
training-summary.json.
Model tree for joshuasundance/myponline-checkpoint
Base model
Qwen/Qwen2.5-1.5B Finetuned
Qwen/Qwen2.5-Coder-1.5B Finetuned
Qwen/Qwen2.5-Coder-1.5B-Instruct