Datasets:
v1.0 reproducibility and independent replication
The public release contains source, synthetic demonstrations, complete numerical receipts, and independent audits. Original checkpoint files, calibration/fresh tokens, teacher/student activations, fitted model arrays, optimization states, and real-model finite wire payloads are excluded. Their recorded hashes remain public. A new replication downloads the exact public inputs and builds a separate experiment. It must never overwrite the published once-only journals.
1. Install the recorded CPU environment
Use Python 3.12 in a fresh environment. From the extracted release directory:
python -m pip install torch==2.14.1+cpu --index-url https://download.pytorch.org/whl/cpu
python -m pip install -r requirements-recovery.txt
The CPU index must remain available when resolving the +cpu Torch package.
For an installer that resolves all requirements together, use:
python -m pip install -r requirements-recovery.txt --extra-index-url https://download.pytorch.org/whl/cpu
The recorded experiment uses NumPy 2.3.5, SciPy 1.17.0, Torch 2.14.1+cpu, Transformers 5.19.0, Datasets 5.1.0, Hugging Face Hub 2.2.0, Safetensors 0.8.0, and Tokenizers 0.23.3. The public requirements pin the main numerical and model packages. CPU architecture and BLAS behavior can still change floating-point results; a replication is a new run with its own source/input receipts.
2. Create an empty, separate experiment and recover inputs
Choose an unused directory outside the extracted release. The
<replica>/working/embrace layout matches the existing experiment entry points,
which locate their cache two directories above the project.
python v1_rebuild_inputs.py --project-root /absolute/path/eve-v1-replica/working/embrace
cd /absolute/path/eve-v1-replica/working/embrace
The helper refuses the original project, its cache root, nonempty destinations,
and nested destinations. It copies public sources and preserves the published
v1 receipts under reference_results/original_v1. New receipts go to results.
It performs no neural forward pass and no quantizer fitting.
The exact download identities are:
| Input | Identity | Pinned revision |
|---|---|---|
| Model and tokenizer | facebook/opt-125m |
27dcfa74d334bc871f3234de431e71c6eeba5dd6 |
| Dataset | Salesforce/wikitext, wikitext-2-raw-v1 |
b08601e04326c79dfdd32d625aee71d232d685c3 |
The helper verifies all seven downloaded model/tokenizer asset lengths and
SHA256 digests, then checks the six original FP32 projection hashes from the
preserved v0.6 protocol. The pytorch_model.bin download is 250,540,281 bytes,
SHA256 2d74da6615135c58cf3cf9ad4cb11e7c613ff9e55fe658a47ab83b6c8d1174a9.
Checkpoint and dataset licensing remain those of their upstream publishers;
this release supplies neither original assets nor a new license for them.
Token preprocessing concatenates each dataset split independently with two
newline characters, then applies the pinned tokenizer with
add_special_tokens=False and no truncation. Contexts reset every 128 tokens.
The calibration input is the first 8,192 train tokens, shaped 64 × 128.
The reserve fit uses contexts 0–47 and its development selection contexts
48–63. All 64 contexts are already used by the base quantizer fit; the reserve
development split is not an independent base-quantizer holdout.
| Token content | SHA256 of contiguous little-endian int64 token bytes |
|---|---|
Original calibration, train 0:8192 |
dd50ee9bd68ba66fcc8c1f81a561429a502c24cfb208efe1df47f2f95ee61689 |
| Entire test split, 287,644 tokens | 98f0f6d425c82678630a55c6f20ba3d873889195e133e6eeb6166bbe5f22f104 |
v1 test window, test 81920:90112 |
4f5293e497155b46c6ec195bab9e73b62e6dfa23a29a840fa15dceda14b1832d |
The original v1 test window is excluded from fitting and role selection. A replication of its 64 contexts measures reproducibility of published scores; it does not create new untouched evidence for the original experiment. Every new experiment has its own once-only evaluation journal.
3. Rebuild teacher evidence and the four base controls
Read the printed model directory, also recorded as model_dir in
results/v1_recovery_environment.json. Substitute that absolute directory for
MODEL_DIRECTORY in these commands:
python v1_teacher_collect.py --model-dir MODEL_DIRECTORY --tokens private/recovery_env/tokens_train.npy --threads 2
python v1_base_build.py --workers 4
python examples/check_v1_base_decode.py
python v1_select_release.py --stage freeze
Teacher collection records all native first-block projection inputs, original weights, second moments, selected top-32 IDs, and direct tail logsumexp. Base building creates binary, source hybrid-fork, one-bit scalar, and two-bit scalar controls for all six complete first-decoder-block projections. These are independent research reference implementations; the scalar controls are not the official GPTQ software. Original base and fork sources remain byte-preserved in the release history. Neither step reads fresh-test labels.
4. Freeze and run both matched nonlinear repair families
python v1_gauge_reserve_fit.py --freeze-only --model-dir MODEL_DIRECTORY --threads 4
python v1_gauge_reserve_fit.py --run
python v1_lowrank_gauge_fit.py --freeze-only --model-dir MODEL_DIRECTORY --threads 4
python v1_lowrank_gauge_fit.py --run
The sparse family has supports of 16 and 256 columns; the dense matched control has ranks 3 and 51. Both use the same three bases, calibration split, terminal 33-category teacher KL, four optimizer epochs, and five decoded finite checkpoints including epoch zero. All 60 checkpoints are retained. Paid FP16 weight and bias reserves, sorted sparse column IDs or both dense factors, all six base projection frames, metadata, checksums, and the complete EEB1 envelope are included in every reported candidate cost. Actual strict decoding followed by FP32 installation defines the scored native model.
The recorded repair protocols use four CPU threads per family. The original run executes both families concurrently, for eight CPU threads in total. The ordered commands above also work sequentially; keeping each family's frozen thread count preserves its declared numerical protocol. Teacher collection uses two threads.
The fitters refuse an existing start journal or result. A start journal remains as the original exclusive execution marker; the separate completed result holds all checkpoint evidence. Do not delete a journal to resume or retune. For a failed replication, preserve the failed directory and start a separately identified replication.
5. Score the fourth bare control and freeze development selection
python v1_select_release.py --stage prepare-gptq1
python v1_select_release.py --stage score-gptq1 --frozen-protocol-sha256 PRINTED_GPTQ1_PROTOCOL_SHA256
python examples/check_v1_release_evidence.py --pre-model --threads 2
python v1_select_release.py --stage propose
python examples/check_v1_release_evidence.py --native-calibration --threads 2
python v1_select_release.py --stage finalize --independent-evidence results/v1_calibration_objective_evidence.json
The catalog contains 60 checkpoints plus four bare incumbents. Checkpoints whose decoded native train KL exceeds their own epoch-zero train KL are retained in the catalog and excluded from role admission. Primary and extreme roles minimize development teacher KL under the actual complete bare two-bit and one-bit EEB1 caps respectively. Ties use complete byte count and stable candidate name. Source hybrid-fork variants are retained under the same caps, including when another family wins globally. The independent native audit replays teacher and selected calibration endpoints without NLL, fresh data, or an optimizer. It checks direct tail geometry and the sharp gauge bridge.
6. Freeze every profile before one fresh evaluation
python v1_model_probe.py --stage prepare
python v1_model_probe_release.py --stage freeze --threads 4
python v1_model_probe_release.py --stage evaluate --frozen-protocol-sha256 PRINTED_MODEL_PROTOCOL_SHA256
python examples/check_v1_release_evidence.py --post-model --threads 2
Use the exact SHA256 printed by the preceding freeze command. The evaluation entry point refuses a preexisting result or start journal. The full native curve, chosen roles, paid bytes, model configuration, source closure, development evidence, and token identities are frozen before the first fresh next-token loss. Exactly identical installed six-weight-plus-bias states share one native score and remain explicitly labeled as aliases. All other 66 projection weights, embeddings, normalization parameters, and remaining biases stay FP32. The reported block rate is not a whole-model sub-two-bit rate.
The release entry point uses a strict canonical JSON comparison for native
configuration metadata. JSON converts Python integer dictionary keys to string
object names; this comparison checks serialization equivalence, rejects key
collisions and nonfinite values, and does not change the model or configuration.
The original run preserves a recorded zero-forward preflight rejection by its
earlier direct-dictionary guard. A clean replica freezes directly with both the
rejected-protocol and failure records absent. It must not fabricate that historical
failure. A partially copied preflight pair is rejected. Original token preparation
through v1_model_probe.py --stage prepare remains valid.
7. Verify the release and build a separate replication report
python finalize_v1_release.py --record-tests
python finalize_v1_release.py --check
The check also requires a current independently recorded manuscript visual-QA
receipt; it does not create a manuscript or claim that a new replication has
the original report. Author and render a report for the new evidence, review
every page, record its current hash in results/V1_PDF_QA.json, then finish:
python finalize_v1_release.py --write-receipt
python package_v1_release.py --output-dir /absolute/path/eve-v1-replica/deliverables
The packager rejects private/model/data leakage, snapshots all public file hashes, creates root indices with explicit self-hash exclusions, and reopens the ZIP to verify its exact member set, CRC, every hash and length, source stability, and standalone/archived PDF identity. It performs no fitting or neural inference. Matching short-window results on this small partial-model study demonstrates replication; broader model, corpus, full-model compression, and compressed-kernel comparisons are required for a SOTA claim.