Terrano09/generalized-cpp-kernel-GRPO20

Generalized C++ kernel GRPO20

Portable PEFT LoRA checkpoints from generalized-cpp-kernel-grpo20-spot-20260829-083214-retry1.

Selected checkpoint

The selected release checkpoint is iter_0000014. Selection basis: higher four-trial Fixed26 Pass@1 and turn-2 means than iter 19.

Field	Value
Base model	zai-org/GLM-4.7-Flash@7dd20894a642a0aa287e9827cb1a1f7f91386b67
Adapter SHA-256	b4bb3a250e28696c597d84db459caa75978e160996818dbfce22b8896b2c794b
LoRA rank / alpha	16 / 32
Target modules	q_a_proj, kv_a_proj_with_mqa, o_proj, gate_proj, up_proj, down_proj
Planned updates	20
Run outcome	Retry completed all 20 updates and preserved checkpoints through iter 19.
Training data	Generalized_CPP_GRPO20_train.jsonl, 6 rows
Training-data SHA-256	566ea43aad4f6419b345e0295288c27b374575ce4db30cbd2d6b7510bf86ffbe
Training-manifest SHA-256	25d6165985c440e0bc79f3a3f2a6aeeac4943324f2289bc03fd1d775d30bf3a6

Post-training evaluations

Each row uses only its selected best four receipt-verified trials (26 tasks per trial, 104 task evaluations). Iterations are reported separately.

Checkpoint	Pass@1 trial scores	Pass@1 mean	Turn-2 trial scores	Turn-2 mean
iter_0000014	13, 13, 11, 10	11.75/26	17, 17, 14, 16	16/26
iter_0000019	12, 10, 10, 9	10.25/26	17, 13, 13, 11	13.5/26

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_0000014	1.50; 10-13; 8-15.5	1.41; 14-17; 11.75-20	17/57 (29.8%; CI 14.5-48.9%)
iter_0000019	1.26; 9-12; 6.25-14.5	2.52; 11-17; 9.25-17.75	13/63 (20.6%; CI 9.5-35.0%)

Evaluation used fixed26-contract-v2, thinking enabled, temperature 0.7, top-p 1.0, and a 32,768-token response limit. The complete selected run IDs and byte-for-byte receipts are under evaluations/.

Training data

6 executable shadow-task rows: clock, complex-numbers, crypto-square, grade-school, kindergarten-garden, and perfect-numbers. The manifest explicitly records overlap with six Fixed26 task IDs; reference answers are not packaged.

The exact JSONL and its source manifest are included at the repository root. Their hashes are checked during release construction.

Checkpoints

Every checkpoint directory contains only the two portable inference artifacts: adapter_config.json and adapter_model.bin. Megatron tensor-parallel shards, optimizer state, and other training-only files are intentionally omitted.

Checkpoint	Adapter SHA-256
iter_0000004	cf25cdb386b4e9cbc372480768882dc9f828b5b68e9017868c0255945d30aec6
iter_0000009	83eec62a120142917902355a641043cf1970658170056b6fce5ebf98dd1de7dd
iter_0000014	b4bb3a250e28696c597d84db459caa75978e160996818dbfce22b8896b2c794b
iter_0000019	e915a82a09a290cd4a6b587fe08fdad8acbbf58dea27e1a53e9e77f9effdf300

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "zai-org/GLM-4.7-Flash"
checkpoint = "Terrano09/generalized-cpp-kernel-GRPO20"
subfolder = "checkpoints/iter_0000014/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 plus four aggregate receipts and eight shard receipts for each reported evaluation row. Checksum files bind each evidence bundle.

These are assisted Fixed26 regression results using selected best-four cohorts, not pristine held-out benchmark claims. The Generalized C++ dataset, where applicable, explicitly overlaps six Fixed26 task IDs; consult its included manifest before comparing results.