Terrano09/midband-ten-kernel12-set2-GRPO30

Midband Ten kernel12 Set2 GRPO30

Portable PEFT LoRA checkpoints from midband-ten-kernel12-set2-grpo30-spot-20260828-143019.

Selected checkpoint

The selected release checkpoint is iter_0000014. Selection basis: best receipt-backed Fixed26 checkpoint among the preserved checkpoints.

Field	Value
Base model	zai-org/GLM-4.7-Flash@7dd20894a642a0aa287e9827cb1a1f7f91386b67
Adapter SHA-256	ecc5b1e459a95032f06f4e10fde431ca28acd9ecfb35180ff6d70baa3305e402
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	30
Run outcome	Spot-preempted after rollout 15; checkpoints through iter 14 were preserved.
Training data	Midband_Ten_Set2_GRPO30_train.jsonl, 80 rows
Training-data SHA-256	36d031392e940514f8d7ce82a01663c7592cf7f505e2d455a159170324686360
Training-manifest SHA-256	e8ec90224ccc2e8b620c63c77e7a9ea01aff35d26863270ff5e5c6da62a899eb

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	14, 11, 10, 10	11.25/26	15, 14, 14, 13	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_0000014	1.89; 10-14; 7.25-15.25	0.82; 13-15; 10-18	11/59 (18.6%; CI 9.6-29.6%)

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

80 Set2 rows: 8 each for allergies, circular-buffer, clock, complex-numbers, grade-school, parallel-letter-frequency, perfect-numbers, phone-number, robot-name, and spiral-matrix.

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	4b81db525690834cce03c4deb79a54a97aa06673a427f549a137b438657e8891
iter_0000009	aa883d245371afa09390f411535baadececde1e387c179155719123e08f078c9
iter_0000014	ecc5b1e459a95032f06f4e10fde431ca28acd9ecfb35180ff6d70baa3305e402

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "zai-org/GLM-4.7-Flash"
checkpoint = "Terrano09/midband-ten-kernel12-set2-GRPO30"
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.