TokenBender/glm47-bank-account-official-grpo20

GLM-4.7 Bank Account Official GRPO20

Complete training and evaluation artifacts for the execution-grounded GRPO run tracked in issue #111.

This repository is public with manual access approval. It contains model adapters, per-rank resume state, generated rollouts, detailed execution evidence, and offline W&B bundles. It does not duplicate the third-party GLM-4.7-Flash base weights or contain access credentials.

Result

Surface	Result
GRPO updates	20/20 completed
Final in-training strict reward	111/256 (43.3594%)
Official bank-account full-solve evaluation	57/64 (89.0625%)
Format or compilation failures in official eval	0/64
Remaining official-eval failures	7 stale-balance-after-reopen failures

The official evaluation is an intentional exact-task overfit measurement, not held-out generalization evidence. The training manifest explicitly declares bank-account as an authorized official target with zero validation rows.

Fixed26 four-trial evaluation

The best saved checkpoint, iter_0000019, was also evaluated with the same fixed26 method used for Synth-v1 epoch 50: four trials, temperature 0.7, top-p 1.0, thinking enabled, and up to one compiler/test-feedback repair turn.

Metric	Trials	Mean
Pass@1	11, 8, 6, 8	8.25/26 (33/104, 31.73%)
Multi-turn with feedback	13, 14, 12, 13	13/26 (52/104, 50.00%)

• Four-trial union: 20/26 tasks
• Bank account Pass@1: 2/4 trials
• Bank account by turn 2: 3/4 trials
• Synth-v1 epoch-50 comparison: Pass@1 9.5 -> 8.25; turn 2 12 -> 13

The complete fixed26 evidence package contains all four merged attempt directories, both shard receipts and logs per trial, task outputs and Aider histories, the suite receipt, exact active evaluator and launch plan, the BKM diff, a machine-readable summary, and repository-wide checksums.

Identity

• Training run: issue111-bank-official-grpo20-20260817T151213Z
• Evaluation run: issue111-bank-official-grpo20-20260817T151213Z-eval64-r1
• Execution source: 2c069b13c61afa860a0d62d5eb0bc73ba83ab933
• Runtime image: ghcr.io/tokenbender/glm47-runtime@sha256:b4f67ba1519bf276fe1dcb6fcb457600e4a256bc0d5c3011fcbd9cc05f240d43
• Base model revision: zai-org/GLM-4.7-Flash@7dd20894a642a0aa287e9827cb1a1f7f91386b67
• Synth-v1 source adapter: TokenBender/glm47-synth-v1-100ep, SHA-256 4acb7f23c295f45380155c5d9ee6bc59422262f0cb51f0c02f7e550d405b575a
• Final GRPO adapter: iter_0000019/adapter/adapter_model.bin, SHA-256 186b0fc5b200fb8bb55bf85ee4416f2682a470580f0231c6f3f2a4d414bd898e
• Training data manifest: SHA-256 b9c80354d4d05123f8a3768898379ff215251fdf3444e6e0f05d52b02b968299
• Official Polyglot source: Aider-AI/polyglot-benchmark@7e0611e77b54e2dea774cdc0aa00cf9f7ed6144f/bank-account
• Hardware: one GCP a3-highgpu-8g Spot node, 8 x H100 80 GB

Training contract

• Eight executable bank-account task states, 32 samples per prompt, 256 samples per update
• Learning rate 3e-5; warmup fraction 0.1; constant schedule
• Sequence length 6144; maximum response length 4096; temperature 0.7
• LoRA rank 16, alpha 32
• Policy and frozen KL reference initialized from the same Synth-v1 source adapter
• low_var_kl loss with coefficient 0.1
• Checkpoints after updates 5, 10, 15, and 20 (iter_0000004, iter_0000009, iter_0000014, iter_0000019)

Each checkpoint contains the HF-format adapter, four Megatron TP adapter shards, and eight per-rank training-state files required for same-stage continuation. The base model and frozen reference checkpoint remain external dependencies and are identified above rather than duplicated here.

Repository layout

runs/issue111-bank-official-grpo20-20260817T151213Z/
  checkpoints/       four adapter and resume-state checkpoints
  data/              exact GRPO, SFT, monitor data, shadow files, and manifest
  grpo_lora_r16/     run log, receipt, gate, manifests, evidence summary, VRAM logs
  rollout_dumps/     all 20 training and monitor rollout dumps
  wandb/             preserved offline W&B bundles

evaluations/issue111-bank-official-grpo20-20260817T151213Z-eval64-r1/
  official_eval64_summary.json
  official_eval64_records.jsonl
  rollout_dumps/     accepted 64-sample dump and monitor dump
  data/              exact evaluation inputs and executable task material
  grpo_lora_r16/     evaluation logs, receipt, manifests, and evidence summary
  wandb/             preserved offline W&B bundles

fixed26-evaluations/issue111-grpo20-iter19-fixed26-mt2-suite-20260817T193037Z/
  SUMMARY.json        compact four-trial result and Synth-v1 comparison
  suite/              authoritative suite receipt
  attempts/           complete merged evidence for trials a1 through a4
  method/             active evaluator, overlay, plan, and BKM diff
  SHA256SUMS          checksums for all 1,452 evidence files

SHA256SUMS           repository-wide file checksums

The W&B material was produced with WANDB_MODE=offline. The raw bundles are included for later inspection or synchronization; no online W&B URL is claimed.

Integrity

The two source GCS prefixes contain 220 objects totaling exactly 3,171,553,773 bytes. The staged upload matched that object count and byte total before repository metadata was added. SHA256SUMS covers every uploaded file except itself.

The training gate passed and verifies all four checkpoints, adapter hashes, native shards, training-state presence, the source commit, source adapter, task count, and data manifest. The accepted official evaluation summary explicitly excludes one startup-monitor record and scores only the declared 64-sample grpo_0.pt dump.

The fixed26 evidence package contains 1,453 files totaling 79,214,476 bytes including its checksum manifest. All remote file sizes and all 1,453 Git blob identities were verified against the local closeout package with zero mismatches. Its SHA256SUMS file has SHA-256 c937b41bcc4095802af8a509a0bceb692f8c2197ddf861509b37d4fcfdbb5dcf.