SourceHat RLCD returns typed categorical, binary, and ordinal decisions from
a shared Clef/Qwen-derived backbone. This edition contains the complete backbone
with all 496 trained text updates mathematically folded into its dense weights,
plus the FP32 joint decision head, in one model.safetensors file.
There are no adapter files, runtime LoRA branches, PEFT dependency, or separate base-model downloads. Tokenizer, configuration, and the inference wrapper are included.
This edition's bounded population-quality check passes. Numerical fidelity remains a separate result; this is a private review draft, not an all-gates-passed declaration.
Bold values are best within each matched population; ties are bold for both models.
| Population / model | Correct fields | Correct objects | Response coverage | Field NLL | Field Brier |
|---|---|---|---|---|---|
| Historical: unmerged reference | 410/439 | 230/256 | 100% | 0.151181 | 0.082997 |
| Historical: Dense INT8 | 411/439 | 231/256 | 100% | 0.150271 | 0.081609 |
| Fresh: unmerged reference | 305/326 | 111/128 | 100% | 0.145430 | 0.077869 |
| Fresh: Dense INT8 | 306/326 | 111/128 | 100% | 0.142312 | 0.075209 |
Both populations have 100% response and probability-scoring coverage. The reference is the trained, unmerged checkpoint, not original Clef or a new live control. These are bounded release checks, not the sealed final campaign or a production-risk guarantee.
The source dense BF16 model's numerical-fidelity result is failed, with the unchanged 0.020000 tolerance. This quantization's separate four-probe maximum probability drift is 0.041194; its diagnostic fidelity flag is false, and all top-1 choices are unchanged. Population-quality success does not erase either fidelity result.
Across the complete historical and fresh populations, this edition's maximum probability changes from the unmerged reference are 0.151532 and 0.105482. The four-probe drift is not presented as the global maximum.
Candidate-order and opaque-ID checks each cover 32 pairs, with maximum logit differences 0 and 0. Descriptions and semantics are unchanged; this does not test arbitrary paraphrases.
The complete weights file is 30,890,995,272 bytes. Peak PyTorch-allocated and reserved memory are 29.58 and 30.64 GiB on one 64-GiB NVIDIA CMP 170HX. This is not a validated 16-GiB-card fit claim.
Mean warm request time is 2.554 seconds over 452 newly executed requests, excluding loading and reused predictions. The bounded worker executes 453 requests including its public example; a separate isolated offline CLI executes that same public example once more.
The standalone CLI passes in a fresh python -I process with private project PYTHONPATH removed. It uses the existing dependencies; a clean dependency reinstallation is not claimed.
All models below scored the same 256 cases and 439 fields with 100% response coverage. These are historical training-checkpoint results. This quantized edition's own measurements and separate fidelity diagnostics are reported in its deployment check.
| Model | Correct fields | Correct complete objects | Field NLL | Field Brier |
|---|---|---|---|---|
| Original full Clef | 395/439 (90.0%) | 216/256 (84.4%) | 0.2519 | 0.1268 |
| Trained head only | 403/439 (91.8%) | 223/256 (87.1%) | 0.2041 | 0.1070 |
| RLCD trained checkpoint | 410/439 (93.4%) | 230/256 (89.8%) | 0.1512 | 0.0830 |
The population contains 160 executable synthetic cases, 48 Banking77 cases, and 48 CLINC-OOS cases with restricted candidate lists. Training uses seed 126. The aggregate supplement distinguishes historical measurements from the current package's verification status.
The decision contract accepts 1-4 text-only fields: categorical choices with 2-16 candidates, binary propositions with true then false, and ordered scales with 2-14 levels. The compiled limit is 2048 tokens; oversized requests are rejected, not truncated. Caller IDs are not model features.
The model proposes supplied choices and raw field distributions. It does not execute actions or independently estimate object correctness, safety, or workflow success. Supply applicable policies and meaningful fallback choices.
hf download SourceHat-Labs/RLCD-27B-INT8 --local-dir ./sourcehat-rlcd
cd sourcehat-rlcd
python -m pip install -r requirements.txt
CUDA_VISIBLE_DEVICES=2 python inference.py --request example_request.json --device cuda:0
Select the physical GPU appropriate to your machine. Use the included typed-head loader, not a generic chat-generation loader. INT8 merged text weights, retained BF16 components, and the FP32 head are packaged in one weights file. Measured runtime and quality results are reported in this edition's deployment check. This is not a validated single-16-GiB-card runtime claim.
The exact base/tokenizer is Cloudflare/clef revision
0b331204bb13fbd2ca93a64df1956f5b55478ce5. Training used supervised CE, Brier,
ordinal RPS, and a reference KL term, not policy-gradient RL. The package is
Apache-2.0; see LICENSE and NOTICE for attribution.
Training-source references: Banking77,
CC BY 4.0; CLINC-OOS, CC BY 3.0.
No training records, scripts, private prompts, logs, or intermediate checkpoints
are included. This is a research model, not a safety or production-risk certificate.
Version 1.0.0 uses the public typed-decision-v1 contract. CPU checks prove that the renamed wrapper preserves native compilation and fixed-logit responses. The BF16 and INT8 containers have branding-only header changes; tensor bytes are unchanged. EXL3 weights are byte-identical. Historical source hashes identify the original measured containers, not a newly executed quantization. The earlier offline CLI checks exercised the original wrappers; no new neural execution or evaluation was performed for this packaging migration. All model repositories and website candidates remain private/review-only.
See the SourceHat RLCD article.
SourceHat-Labs/RLCD-27B