Cam1

Cam1 is a custom decision model based on the pinned Qwen3.5-4B backbone, a LoRA adapter, and a joint-schema decision head. It returns Choice probabilities and Noul decisions directly. This repository contains the model assets and the minimal inference runtime needed to use them. The upstream backbone is downloaded separately at the revision in inference_config.json.

The adapter, decision head and tokenizer files are byte-identical to the evaluated checkpoint. The export removes training-only metadata and optimizer state, so its packaging hash differs. It requires the custom decision-model inference runtime; the adapter alone is not the complete model. The validated serving policy is full FP32, reference kernels, GPU placement and a 16,384-token context limit.

The completed official Decision Index 0.3 public suite scored 30.72 Public points, with complete: true, 140,178 scoreable requests across 43 benchmarks, and 138,071 successful responses. The 2,107 scoreable refusals comprise 107 overlength requests and 2,000 PhishNChips inputs rejected by the evaluated schema validator. No evidence or options were truncated. See the submission results. Public points are chance-corrected, coverage-adjusted index points, not accuracy. This is not a maintainer Full score or leaderboard admission claim.

cam1_engine:Cam1Engine is the default evaluated behavior. The separate cam1_engine:Cam1CompatibilityEngine accepts empty Noul criteria and null Choice descriptions. That generic post-evaluation fix answered all 2,000 Phish inputs with 51.2% verdict accuracy and exactly preserved the 507-request regression set; it has no new complete-suite score. Its results are not merged into the submitted run. Both entrypoints use the same weights, tokenizer, prompts, FP32 numerical policy and temperature.

Reproduce inference

Use Linux, Python 3.12 and one NVIDIA RTX PRO 6000 with 96 GB VRAM. The evaluated system used Python 3.12.15, driver 580.178.04, Torch 2.11.0/CUDA 13.0, Transformers 5.10.2 and PEFT 0.18.1. The original container was python@sha256:edd0b3ec946cc68bd20e39480bd03929016d0fb6254cbf870cd3bc55ea25116d. Follow REPRODUCE.md for the pinned setup and runner command. runtime.zip contains only inference source, its dependency lock and release identity binding; no training scripts, optimizer state, planning notes or evaluation questions are included.

The runtime requires a fresh process per schema policy, full FP32 parameters, GPU-only placement, math SDPA/reference kernels, disabled TF32/autocast, deterministic algorithms, and a 16,384-token limit. Temperature is fixed at 1.333521432163324. It was originally calibrated under BF16; calibration validity under this FP32 runtime is unproven. Cross-runtime equivalence to BF16 is not claimed. Score-primitive quality is not established by the Choice/Noul evaluation.

Measured successful-request latency in the complete run was mean 408.97 ms, median 255.28 ms, p80 398.75 ms and p95 915.28 ms. These public-suite measurements do not certify the maintainers' separate held-out latency test.

The joint-schema head is derived from Cloudflare's Apache-2.0 implementation (revision 17f0b0ad64efb65d273590632833508766b2aae6). Qwen3.5-4B is attributed to the Qwen team. See LICENSE.

Architecture and what was trained

Cam1 is a trained custom decision model, not an untrained Qwen checkpoint. Its Qwen3.5-4B backbone is augmented with a learned joint-schema head and a learned rank-eight LoRA adapter. The head consumes hidden states to produce typed distributions directly; ordinary autoregressive answer-text generation is not the inference interface.

The joint-schema head has input hidden size 2,560, width 1,024, two routing layers, four layers, 16 attention heads, feedforward width 4,096, and dropout zero. The frozen-backbone head stage trained 112,322,564 parameters. The subsequent head-plus-LoRA stage trained 128,555,012 parameters. LoRA rank is 8, alpha 16, dropout zero; the exact target-module list is in checkpoint/adapter/adapter_config.json. The head architecture derives from Cloudflare's Apache-2.0 Clef implementation at 17f0b0ad64efb65d273590632833508766b2aae6; the trained parameters are specific to this experiment.

Training supports Choice (categorical), Noul (boolean), and Score (ordered-level distribution) judgments. The official public release evaluation below concerns Choice/Noul and does not establish broad Score generalization. This repository requires the custom runtime; loading only its LoRA adapter with AutoModelForCausalLM does not reproduce Cam1.

Training data and recipe

The actual training set is published at Cam1-training-data, with byte-exact original files, a lossless Parquet representation, hashes, source revisions, and source-specific license attribution.

Source	Training records
Stanford SNLI	8,000
PolyAI BANKING77	5,000
Google Research GoEmotions	5,000
LocalLLaMA Typed Decisions	960
Controlled synthetic decisions	1,040
Total	20,000

The original pilot used separate validation (565), calibration (1,006), and evaluated test (11,028) records. Source/document grouping and exact/lexical near-duplicate checks preceded schema augmentation. Question/option identifiers, order, candidate subsets, and semantic wording varied while labels stayed aligned. Neither the original test nor the official public benchmark was used to select the checkpoint or fit this release's temperature.

Training ran locally on an RTX 4070 SUPER (12 GB), using Python 3.12.3, PyTorch 2.11/CUDA 13.0, Transformers 5.10.2, and PEFT 0.18.1. The backbone ran in BF16; head master parameters were FP32 under BF16 autocast. Native training used the original SDPA path, an 8-bit optimizer, batch one with eight-step gradient accumulation, seed 3407, and a 2,048-token training limit. Training kernels were not set to deterministic mode; seed identity alone does not guarantee bit-identical retraining.

Stage	Actual recipe
Frozen-backbone full head	Three epochs, learning rate 0.0003, 7,500 optimizer steps; validation-best checkpoint at step 4,000
Selected native LoRA	One epoch, 2,500 optimizer steps; head learning rate 0.0001, backbone LoRA learning rate 0.00002; rank 8
Distillation	Fixed 2,000-record training subset labeled by the local Clef teacher; distillation weight 0.25; original source labels retained
Selection	Lowest balanced validation reference cross-entropy across declared candidates; selected lora-kd025-v2/step-0002500, loss 0.48644513644996396
Calibration	Independent 1,006-record cohort; fixed temperature 1.333521432163324, carried unchanged into the later FP32 release

Full/compact heads, hard-label LoRA, and a one-epoch generated SFT baseline were compared. The selected native checkpoint was chosen on validation before testing. The original training file SHA-256 is b46db1dd4a0edd2d50e8a43345e5d67e0db66b7b30972bdac7506b4302fbb736; its actual teacher-augmented LoRA input is c5f2c6dcf83221101e9c5d99d94bb849e2c929ae2afd9b67532497883a35797d. training-provenance.json records the key settings and artifact identities.

Original pilot results, distinct from the official public score

On the fixed original evaluated cohort of 11,028 records and 18,846 questions:

Model/interface	Question accuracy	Schema-valid responses
Cam1 selected native model	79.15%	100%
Generated SFT baseline	77.75%	98.14%
Unmodified generated base	51.96%	68.69%

Invalid generated outputs count as wrong. Native minus base is +27.19 percentage points, with paired 95% group-bootstrap interval [+25.82, +28.53]; the original pilot promotion gate passed. Native minus SFT is +1.40 points [+0.80, +1.99], with a 0.84-point Noul regression. These are end-to-end label-plus-format results, not isolated reasoning gains. The separate sampled public 0.2.1 comparison likewise does not establish a general reasoning advantage: its common-valid native-minus-SFT interval includes zero.

The complete official Public Index 0.3 result remains 30.72, as described above and in the linked results dataset. The BF16 Windows pilot, sampled offloaded evaluations, and FP32 Linux release evaluation use different numerical/runtime conditions; their latency and probability results must not be conflated.

Intended use and limitations

Cam1 is an experimental text decision interface for explicit criteria and bounded candidate choices. It is not a general chat model, an authorization authority, or a substitute for validation in a new domain. Confidence is a distribution statistic, not proof of correctness or permission to execute an action. Source annotations and synthetic teacher labels can be noisy or biased; unknown pretraining contamination and templated-case correlations remain. Long-context stress probes do not establish general long-document quality. The release rejects unsupported complete inputs rather than silently truncating evidence.

This model card documents the final retained original Cam1. The closeout updates documentation and publishes its original training data; it does not replace the adapter, head, tokenizer, runtime, or frozen evaluation outputs. The original evaluated release remains reproducible at revision 777259d8384497d51da693115587ddacc515f4d6.