Terrano09/midband-ten-set2-iter4-GRPO40

Midband Ten Set2 iter4 GRPO40

Portable PEFT LoRA checkpoints from midband-ten-set2-iter4-grpo40-spot-20260831-205120.

Selected checkpoint

The selected release checkpoint is iter_0000025. Selection basis: best training-monitor checkpoint (Pass@1 0.71484375; Pass@2 0.8371975806451613); iter 39 was evaluated but is not labeled best.

Field	Value
Base model	zai-org/GLM-4.7-Flash@7dd20894a642a0aa287e9827cb1a1f7f91386b67
Adapter SHA-256	fc27e6ec9bfeef5c33c7769cef5a1fb62d65dabb85c9bd8efb4a0c0daa5ec400
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	40
Run outcome	Completed all 40 updates, warm-started from iter 4 of midband-ten-set2-stable-grpo40-spot-20260831-133241.
Training data	Midband_Ten_Set2_GRPO40_train.jsonl, 80 rows
Training-data SHA-256	c5f4516933a733c0b63edc7b2bcf35afc1367bb788adefa514fee5ca832ca4a9
Training-manifest SHA-256	820086e7e749d38a3e82f6fa3971b544929bc6f315d448265753d2fcc3df2df2

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_0000019	9, 9, 9, 9	9/26	13, 14, 11, 9	11.75/26

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	0.00; 9-9; 5.75-12.5	2.22; 9-14; 8.25-15.25	11/68 (16.2%; CI 8.3-25.4%)

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/.

Secondary single-trial checks

These checks are disclosed separately and are not mixed into the four-trial iter-19 aggregate.

Checkpoint	Run ID	Pass@1	Turn 2
iter_0000025	midband-set2-iter25-fixed26-contractv2-thinking-20260901T211642Z	8/26	12/26
iter_0000039	midband-set2-iter39-fixed26-contractv2-thinking-20260901T211642Z	8/26	10/26

The iter-39 check does not change the selection label: iter_0000025 remains the best checkpoint by the training-monitor criterion above.

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_0000000	80482813aae92ca46bd3d86b0c9c10ec0acf8fab313d5186111805be21299f3f
iter_0000001	27cc4636520839d8a9fe2a068127ca088eef431127230ca422cd9bcf94145887
iter_0000002	1a336810d9a0f11bd3846113cca8ea3cc2e3033b02fd44d580cd9ddfe20cbe82
iter_0000003	9f22fcf44b3f8f76655e2f747d0ef73dbcd6672ec4ac126c37dfd9b334fdb2ff
iter_0000004	6e1ec5968871af2968d47cf5f1c39e68f00784e520e32d98a642dca5d046ce15
iter_0000005	823c93c50c90b52f20797246c06f65b5f2d908e1296449c5406f9e4dcddc6612
iter_0000006	472a47f8e4e87192690bd58a21170afdea89b570cf75d457a4055d5f439c04d1
iter_0000007	f24c56aaa8ae3a8a18beeb8aea928d2b43631d466ffe997d4cf2529e9979e667
iter_0000008	4dc1db616ce9d110f9435f3413da2724e49cae723bc038978cca224ff1ad187a
iter_0000009	fd2680ea3a3fd22296acd1cba81e1f523f39943b4c65e9dbdc47c682784083d5
iter_0000010	3c4742b57f9f23fe981dd3161641a2946b9bbccf7c4a47bf006364adcd6c4fc5
iter_0000011	b74422ca9e92a213c54eb526a9ac42eeba99f09658c7510264b64c07ca2a8d9b
iter_0000012	360861e9ee0ba5354aaa1f86153dacdc7dfc2344b0d200f2c824df2f66200c60
iter_0000013	e5073fdfa22e0499dfd818be5922f2528aaef369b2d92148b2a75a22327afa74
iter_0000014	1efc69bf2ebcaf37790ba1b701d198b99381a26b54b8093e3c454a2ac504515b
iter_0000015	7aebdd0b7bc9b52d377c505a5168754d12bcca55e9ab5c32e38e48afcb055a3a
iter_0000016	8fcd888a2f18c99f26179b9604fae8b5d84fe7cb06597a3d1d7809ed229b5066
iter_0000017	ae3475dda0659e0f5b111533e151041a94b229838bf823a5fe5607570dfa008f
iter_0000018	261d138dd2b9699ad2da989687ac349103c0eab75836e348559aaadbd9160270
iter_0000019	987c7ebb0c81e0df3f02539a237bf8cae1e68add96843917b396aa276f98d1ed
iter_0000020	fb3bacd02933a37b330f08edc728891880c2ee20f6199a98b6e032053b0dc086
iter_0000021	4d2405639e25fead59ff2ea691e706e6dbf760a2b47698eda80278b59525a036
iter_0000022	b7d4ee5a82bd739907359d57906062f9006ed4ba885e908fe09e044b58611bac
iter_0000023	9b12fb0ab1431581648503ff8b7d1731daf14bdb696d6f7a18b58cfdfe93ffbd
iter_0000024	e5f5640fa116ab3717fd4f02e46fea0461bea3cb93f3b42c272a0aff311096b0
iter_0000025	fc27e6ec9bfeef5c33c7769cef5a1fb62d65dabb85c9bd8efb4a0c0daa5ec400
iter_0000026	8f0f530012bb1bd845556bd6a2706bd83ec82cebf68723905e43d314ae7f66de
iter_0000027	e45991c97f4778278be08d0e58d3e16db5939ad246e1945cc145cc5f58648d09
iter_0000028	15c66612d3a4386f930286a2bba8192ce3f3f7b9e23036a32e81de111a6d144f
iter_0000029	780e211fbf6e96905a7ab42e69ad5e23a09749e833336062e491fd1e770d9acc
iter_0000030	7985f595d2bca0c29e5ea8a3a865bac29e74419d9f3b057c81c3c8d6a32ae52d
iter_0000031	911d1ce9841e66ab25dd5bf33bb06794290e006b08bb4c1bc21c2cfa95e7ed05
iter_0000032	80ad5da2edb9278f8f26cd8dff93d9d63369ebd545f57a9559995d8f8eae91f5
iter_0000033	2d1365a8e14ff98c1e71d46e01885514687a3f6331473ac5a413bb22f866e013
iter_0000034	c9866088ce5f0c5f5aebfeb1495a9e37835fee1c1637e6108f2d843f7f415ab8
iter_0000035	3a551de5b08695b9289c0fc9b04069336369334eb7fa7c51df2192f0f810e138
iter_0000036	e8219924e23771a59b90ca30e6be488c1999a0009143c5cdd9aa13c192901db1
iter_0000037	20c87ea9031c3116c5fd1c94ec2469baf08888cda1ccf2d2bf12c940571e787c
iter_0000038	acb06f485146937a87b606f1d372fd1d680528f21243105262da754e5193399a
iter_0000039	17a0343c9c73f725db2073cce02339f4f7f0de012be6f24bfd5e551f1abe8744

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "zai-org/GLM-4.7-Flash"
checkpoint = "Terrano09/midband-ten-set2-iter4-GRPO40"
subfolder = "checkpoints/iter_0000025/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.