Generalized C++ GRPO20 Sept 1 — iter 19
Portable PEFT LoRA checkpoints from generalized-cpp-kernel-grpo20-spot-20260901-065545.
Selected checkpoint
The selected release checkpoint is iter_0000019 (latest preserved checkpoint of the run; 20/20 rollouts completed, training gate passed).
| Field | Value |
|---|---|
| Base model | zai-org/GLM-4.7-Flash@7dd20894a642a0aa287e9827cb1a1f7f91386b67 |
| Adapter SHA-256 | 0c0da7a1c3553dfda66af622c8eb5002f3d7418321cfda261414ef020353240c |
| Adapter config SHA-256 | 0bd6d85f88fc42fefa52627b3c261f1ad58bb2c9519332ae8034dd5dffe2498e |
| LoRA rank / alpha | 16 / 32 |
| Warm start | phone-number-grpo20-iter14 hybrid adapter (source SHA-256 62fa190ad26e30fc1b5dd9543936ef549a49dd8cfa8220e4af726a1d499e575a) |
| Source commit | 08be9518d7b3609312c51c686b139d011d802d21 |
| Planned updates | 20 (completed 20/20 rollouts; wall 10,771 s) |
| Training data | Generalized_Cpp_GRPO20_Sept1_train.jsonl, 6 rows |
| Training-data SHA-256 | 566ea43aad4f6419b345e0295288c27b374575ce4db30cbd2d6b7510bf86ffbe |
| Training-manifest SHA-256 | 94803063131c19cbf6f89698e350bab5e0f4a7d2ffa87ba0a06077e186b668bf |
| Runtime image | us-central1-docker.pkg.dev/lifeandhalf-24122025/w8-biayn/glm47-full-v5-unadmitted-grpo@sha256:5df1c41efb6062a292ac68a49d68ad955d9f8209c182d3fc2f7e8c634ca076a2 |
| W&B | glm47-generalized-cpp-grpo/generalized-cpp-kernel-grpo20-spot-20260901-065545 |
Post-training evaluations
Fixed26 (fixed26-contract-v2), thinking enabled, temperature 0.7, top-p 1.0, 32,768-token response limit, tries=2.
All four planned trials completed (t01-t04) and every receipt passes the integrity gate
(status=complete, 26 terminal tasks, 26 unique testcases, tries=2). t04 required one shard resume;
per-task outcomes reproduce its aggregate receipt exactly.
| Checkpoint | Pass@1 trial scores | Pass@1 mean (4 trials) | Turn-2 trial scores | Turn-2 mean (4 trials) |
|---|---|---|---|---|
iter_0000019 |
12, 8, 9, 11 | 10/26 | 13, 15, 13, 15 | 14/26 |
| Checkpoint | Pass@1 SD; range; task-bootstrap 95% CI (out of 26) | Turn-2 SD; range; task-bootstrap 95% CI (out of 26) | Conditional turn-2 recovery |
|---|---|---|---|
iter_0000019 |
1.83; 8-12; 6.25-13.75 | 1.15; 13-15; 9.75-18.25 | 16/64 (25%; CI 13.16-40.38%) |
The matched Phone Number kernel12 GRPO20 iter_0000019 comparison cohort (required by RELEASE_PLAN.md) has not been evaluated yet and is pending.
Receipts, shard receipts, per-task results, transcripts, and logs are under evaluations/.
Training data
6 generalized-cpp-v1 rows: clock, complex-numbers, crypto-square, grade-school, kindergarten-garden, perfect-numbers.
All six task IDs explicitly overlap the Fixed26 evaluation tasks (authorized training target per the data
manifest split contract); consult training_data_manifest.json before comparing results.
Checkpoints
Every checkpoint directory contains only the two portable inference artifacts: adapter_config.json and
adapter_model.bin (9,741 tensors). Megatron tensor-parallel shards, optimizer state, and other
training-only files are intentionally omitted. Note the portable adapter includes 207 layer-47 tensors;
the serving adapter used in evaluation strips them (serving SHA-256 6a7b3247245ab652fca0dbf4e628de05e5e636c17770fcb34f5f3ffb9961d4d5).
| Checkpoint | Adapter SHA-256 |
|---|---|
iter_0000004 |
3f041b0e5d7d153cc3e14ee646dd2c476662baec4e34bde30eeeda597aa23808 |
iter_0000009 |
a1267c25b00b7599ae892abdbc6c89ef872285dd10d44e7c6ff27c8b0f47ffa8 |
iter_0000014 |
ada21f6499d117ae9a1ab7d18eb2089ac752976026bc95c4f6aee057ace8216e |
iter_0000019 |
0c0da7a1c3553dfda66af622c8eb5002f3d7418321cfda261414ef020353240c |
Loading
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = "zai-org/GLM-4.7-Flash"
checkpoint = "HimanshuPathak/grpo_iter19_sept_1"
subfolder = "checkpoints/iter_0000019/adapter"
tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True)
model = PeftModel.from_pretrained(model, checkpoint, subfolder=subfolder)
Reproduction and evidence
The release includes the exact training JSONL and manifest, the training gate, checkpoint/artifact
manifests, launch configs, and per-trial aggregate + shard receipts with per-task results and transcripts.
Checksum files bind each evidence bundle (evaluations/.../checksums.sha256).
These are assisted Fixed26 regression results on a complete four-trial cohort, not pristine held-out benchmark claims. The training data explicitly overlaps six Fixed26 task IDs; see above.
- Downloads last month
- 30