28,192-token OpenCode/Hermes follow-up
The paired 28,192-token rerun records base 3/4 and released MLP 3/4 actual client passes. All client requests are fully attributed: True. The same model weights are used. This is a new compatibility profile; historical HE9 scores remain attached to their original 8,192-token identity. No new complete HE9 qualification or promotion is claimed.
The installed Hermes one-shot path omits output-token fields. The frontend now
uses its configured serving allowance when those fields are absent, replacing
the silent 1,024-token default. Both arms use a 28,192 cap/default, full 262,144
context, Q8 K/V cache, unchanged native runtime and instruct sampler. Explicit
smaller client/development requests remain honored. Native generation settings
are audited for every call, including auxiliary client calls. All main Hermes
tool calls use the requested 28,192 allowance.
Table with columns: Fixture, Client, Base, Released MLP| Fixture | Client | Base | Released MLP |
|---|
| File edit | opencode | PASS (23.621s) | PASS (27.133s) |
| File edit | hermes | PASS (63.654s) | PASS (60.117s) |
| Two-module project | opencode | PASS (502.853s) | PASS (597.922s) |
| Two-module project | hermes | FAIL (225.061s) | FAIL (210.825s) |
Development regressions versus the new base: ['bounded_chunks'].
The existing file/project deadlines, approval policy, fixtures and independent
Docker grading remain fixed. Budget delivery is separate from program correctness;
a failed assertion or incomplete attribution remains a failed/invalid trial.
Single observed timings do not establish repeatable latency improvement.
The source fix is 4feadcc8216c7953e6a4e2c2643457be0029bfc6; the full .54 suite passes 243 tests, zero skips
in 46.833 seconds, with pinned Docker. See the complete follow-up,
raw paired report and the new
source archive under code/. The earlier 18 truncated continuation calls remain
in the historical evidence. No model weights, training or provider defaults change.
The full follow-up includes exact .54 commands with --max-output 28192,
authenticated conversion manifests and fresh output directories.
Contents
adapter_model.safetensors: unchanged measured PEFT MLP factors, rank 8,
alpha 16, 96 language-MLP modules / 192 FP32 factors.
adapter_config.json: portable base locator and pinned revision; all other
adapter settings equal the measured configuration.
native/adapter-F32.gguf: unchanged unmerged native LoRA export, 36,189,792 bytes.
METHOD.md, EVALUATION.md, CORPUS-ATTRIBUTION.md: recipe, controls,
failure analysis and upstream credits.
evaluation/: complete per-task results, development/client reports and
qualification decisions, including rejected experiments.
provenance/: original measured configuration, training/export records,
pinned recipe and upstream license.
code/: immutable tested source and the continuation/finalization patch.
Maintained source: groxaxo/weight-space-programming.
artifact-provenance.json, SHA256SUMS: local integrity checks for release files.
The base checkpoint, private grading inputs, canonical answers, credentials,
personal client state and optimizer pickles are not included.
Load the PEFT adapter
Use the language-capable Qwen3.5 conditional-generation class. The official base
is pinned to 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. The measured training environment used PyTorch 2.11.0,
Transformers 5.22.1, PEFT 0.21.2, bitsandbytes 0.49.2 and FLA/fla-core 0.6.
Install a compatible PyTorch/CUDA stack for your hardware before these libraries.
import torchfrom transformers import AutoTokenizer, Qwen3_5ForConditionalGenerationfrom peft import PeftModel base_id = "Qwen/Qwen3.5-4B"base_revision = "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_revision)base = Qwen3_5ForConditionalGeneration.from_pretrained( base_id, revision=base_revision, torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, "groxaxo/Qwen3.5-4B-WSP-HE9-MLP").eval()messages = [{"role": "user", "content": "Write a Python function that returns the unique items in their original order."}]text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)inputs = tokenizer(text, return_tensors="pt").to(model.device)with torch.inference_mode(): output = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.8, top_k=20)print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))
This example shows the loading interface; the published scores use the native
Q8 profile below, not this newly configured Python generation call. Hardware,
precision, sampler, parser and client changes require their own evaluation.
PEFT may round factors through the base layer's BF16 precision while loading.
The separate exact-loading control uses FP32 target placeholders; the measured
native profile retains the unchanged F32 GGUF factors.
Native OpenCode and Hermes use
The original HE9 measured profile uses the official base converted to Q8_0 plus the separate
F32 LoRA. It uses 262,144-token context, 253,952 input / 8,192 output reservation,
Q8 K and V cache, one slot, flash attention, no context shifting and non-thinking
mode. Full allocation and a 202-message request are verified; full-window recall
and coding quality at that length are not established. Training length is 1,536
tokens and is a different limit.
Use the included project's authenticated conversion/export workflow and
scripts/serve_coder_gguf.py frontend for OpenAI-compatible chat/tool parsing.
examples/opencode.json and examples/hermes-config.yaml show isolated provider
profiles. ops/run54_gguf_phase.py is the measured .54 supervisor, with explicit
GPU admission and Bonsai restoration. It contains host-specific policy; inspect
and adapt paths/service ownership before using it elsewhere. The original
conversion manifests authenticate their measured local paths; copying those
manifests alone does not create a portable installation. Direct arbitrary
llama.cpp chat/tool compatibility is not established by this experiment.
The native base SHA-256 is
2164b7e8728dc954702d86fe423da3a240964877d3d31b8f6530d38702c38547.
Adapter hashes and configuration normalization are recorded in
artifact-provenance.json. The base is not bundled. No existing provider default
is changed by this release.
Limitations and next changes
The omitted-limit budget defect is fixed and both arms are retested at 28,192.
See the follow-up results and remaining quality failures above.
Repair bounded_chunks and generalize half-open interval and truthful tool
failure handling; address Hermes permitted-command planning and cross-module
logic. Preserve API validation and approval policy. Add task-family-disjoint
validation that predicts real-client quality and repeat paired trials across
seeds. Longer or lower-loss training alone is not supported as a remedy by the
negative experiments. Run new unseen repository tasks before any promotion.
The original MLP's partial task-76 result regresses from 146 to 47 additional
cases before the first failure. Task 132 changes from finished-but-wrong to
timed-out-and-wrong. Task 32 times out and fails. These remain failures even
though aggregate full-task wins improve. The frozen grader is
reconstructed-base-plus-v1; it is not certified official EvalPlus parity.
Its canonical-root check can reject alternative roots. No new all-164 model
evaluation is performed in this HE9 experiment, and benchmark contamination
cannot be ruled out.
License and attribution
The adapter and project contributions are distributed under Apache-2.0, with
upstream notices retained. The official Qwen base retains its Apache-2.0 terms.
The corpus includes CC-BY-4.0 NVIDIA examples and Apache-2.0 Nous Research examples;
see CORPUS-ATTRIBUTION.md for exact revisions and modifications. Private tests
are not redistributed. This release is independent of Qwen, NVIDIA and Nous Research.