The complete typed-decision model is in one model.safetensors: genuinely merged
text weights quantized with EXL3, the original FP32 decision head, and the lexical
features that head needs. There are no LoRA tensors, runtime adapters, PEFT
dependencies, or separate base-model downloads.
Text projections target 3 bits per weight using EXL3's calibrated allocation.
This covers 400 main attention/MLP projections; the 96 small DeltaNet
in_proj_a and in_proj_b gate matrices remain FP16.
Input embeddings use 8-bit trellis encoding. The lexical output table stays BF16,
and all 122 typed-head tensors stay FP32. Norms, convolutions, and recurrent
parameters use the converter's retained precision. The unused vision tower is
excluded from this text-only edition. Target projection bitrate is not the
average bitrate of every tensor in the file.
The lexical table is mmap-backed in host RAM; only the rows needed for the supplied candidate descriptions move to VRAM. The loader returns every final normalized token representation to the typed head instead of generating chat tokens. Requests do not share a recurrent or KV cache.
The dense BF16 source passes the bounded population-quality checks but not its separate 0.02 numerical-fidelity criterion against the unmerged training model. This edition does not claim exact checkpoint replay. Its own measured checks are reported here.
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 EXL3 3bpw | 410/439 | 230/256 | 100% | 0.149511 | 0.081798 |
| Fresh: unmerged reference | 305/326 | 111/128 | 100% | 0.145430 | 0.077869 |
| Fresh: Dense EXL3 3bpw | 305/326 | 111/128 | 100% | 0.142071 | 0.076329 |
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 0.063988; 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 0.093913 and 0.105065. 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 13,525,326,392 bytes. Peak PyTorch-allocated and reserved memory are 9.88 and 9.91 GiB on one 64-GiB NVIDIA CMP 170HX. This is not a validated 16-GiB-card fit claim.
Mean warm request time is 1.830 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.
The loaded EXL3 source reports version 1.6.4; installed distribution metadata separately reports 1.5.3. The loader checks the loaded source version. The pinned source revision and precision layout are in quantization_config.json.
These are the established 256-case, 439-field training-checkpoint results, with 100% response coverage. They are not measurements of this quantization; EXL3 quality, memory, and unresolved fidelity results are reported in this edition's 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 combines 160 executable synthetic cases, 48 Banking77 cases, and 48 CLINC-OOS cases with restricted candidate lists. See the aggregate supplement.
The decision contract accepts 1-4 text-only fields: categorical choices with 2-16 candidates, binary propositions, and ordered scales with 2-14 levels. Compiled requests are limited to 2048 tokens; longer inputs are rejected, never silently truncated. The caller supplies all choices and policies. Opaque caller IDs do not become model features. Ordinal ordering is semantic.
Use the bundled typed-decision loader, not a generic text-generation server. Install PyTorch first, then the pinned EXL3 source and remaining dependencies. Building EXL3 from source requires a compatible CUDA toolkit and C++ compiler.
hf download SourceHat-Labs/RLCD-27B-EXL3-3bpw --local-dir ./sourcehat-rlcd-exl3
cd sourcehat-rlcd-exl3
python -m pip install torch==2.14.0 ninja
python -m pip install --no-build-isolation -r requirements.txt
CUDA_VISIBLE_DEVICES=0 python inference.py --request example_request.json --device cuda:0
Select the appropriate physical GPU on your own machine. This is the edition intended for 16-GiB deployment; actual memory, hardware compatibility, and probability-quality checks must pass before a validated-fit claim is made.
Outputs include raw field distributions and a proposed object. Object correctness, action success, safety, and workflow success are not independently measured by this checkpoint and remain null. No autonomous action policy or execution is included.
The base/tokenizer is Cloudflare/clef revision
0b331204bb13fbd2ca93a64df1956f5b55478ce5. All 496 trained text updates were
mathematically merged before quantization. This package contains no training data,
training scripts, private prompts, logs, or intermediate checkpoints. Supervised
training used CE, Brier, ordinal RPS, and a reference KL term, not policy-gradient RL.
The model artifact and bundled wrapper are Apache-2.0; see LICENSE and NOTICE. EXL3 is a separate runtime dependency under its own license. Banking77 is CC BY 4.0; CLINC-OOS is CC BY 3.0. This research release does not certify safety or production-domain risk.
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