EVIDENCELOCK — historical research adapters

This repository preserves ten unique historical adapter checkpoints from feasibility checks, controlled SFT variants, DPO experiments and a DPO reference. The retained full-response and state-only adapters are published separately. This archive makes experimental lineage inspectable; it does not designate all checkpoints as independently trained, improved or selected models.

Choose an explicit adapters/<id>/ subfolder when loading a historical adapter. There is intentionally no root adapter configuration. Each subfolder includes byte-identical original weight deltas, a portable normalized configuration, the original configuration and a metadata-change proof. The immutable source archive/member and SHA256 identify its historical provenance.

Adapter subfolder	Recorded artifact role	Weight SHA256
`historical-controlled-attempt`	historical_training_attempt	`78c8018037f7c33c728353f5795a6ed7d1b01affa1f29ec8243d2bd5216a5c52`
`balanced-dpo`	historical_training_candidate	`0b44a8285fee0dd9fe439f23d35d49e932d3eb2bd14b49fbf98da05bce1f4339`
`target-control-sft-and-dpo-reference`	historical_training_candidate_reused_as_dpo_reference	`0396d10fda045b11c701cefea40e0f7568cfaedd305a27edc0c53c502c5bda55`
`dpo-smoke`	training_smoke_adapter	`3a679f0654fc9bc4c756c8453188bf71823869be69c2d8bddc9eee4ff4512218`
`dpo-smoke-reference`	reference_checkpoint_independent_training_not_asserted	`2bce9a3c061879ac8dbc6d797f8ba5ea818ba97cae7c691f3868032be4bb4ad9`
`balanced-sft`	historical_training_candidate	`1ffd274066313c30b0c9e88fabbc4a5bfe72f38d82babcb0ab13f9f34b404225`
`aligntune-smoke`	feasibility_smoke_adapter	`dd3e649c755260733eb90255a6dba611d1d2c846e2454cfebb2a5045e95760b1`
`contrastive-sft`	historical_training_candidate	`c7f74c8addc2b2b34db5bfe4c074bc35c98958fe765ed04a8c0cd2edc2dac3cb`
`dpo-experiment`	historical_training_candidate	`31ba0a424ec788a663f9f9d42c1a026efb75ced25c5a50f6fe766ada654530fd`
`learning-rate-control-sft`	historical_training_candidate	`b1203692eb07c767ac5d55963e43973162fe42a65fe4dc9ae8bb087c60bf1531`

Outcomes and available trainer/evaluation reports are preserved in the research-records dataset; no quality score is inferred from a file's existence. The DPO smoke reference is explicitly a reference checkpoint, with no assertion of independent training. The historical controlled attempt remains an attempt record, with its outcome in the original report. Target-control weights were also used as a DPO reference, so their duplicate occurrences are deduplicated by hash.

Historical experiments reused a small synthetic development split. Frozen final measurements belong to the two separately released retained adapters; these ablations were not selected using those held-out outcomes. All adaptations use the pinned Tiny Aya Fire base; no merged base or tokenizer assets are included.

Attribution and artifact terms

Upstream model creator: Cohere Labs / Cohere. Base: CohereLabs/tiny-aya-fire, revision 73500c80fcb6c8bea3e830c21d5d2920249a0e1a. EVIDENCELOCK contributors produced the LoRA adaptations. The original adapter weight bytes are preserved; only the public configuration's base identifier and revision metadata were normalized. Original configurations and exact before/after changes accompany each artifact.

These adapter deltas are distributed for noncommercial research under CC-BY-NC 4.0 with Cohere's acceptable-use addendum and the Cohere Labs Acceptable Use Policy. Retain attribution and modification notices. Attribution does not imply endorsement. Base weights and tokenizer assets are not redistributed; obtain the base separately under its access conditions. Original EVIDENCELOCK source has a separate MIT license.

Reproduction and records

The final Linux training and evaluation stack is Python 3.12.16, torch 2.11.0+cu130, Transformers 5.22.1, PEFT 0.21.2, TRL 1.7.1, Accelerate 1.15.0, bitsandbytes 0.50.2, CuratorKIT 0.1.22, AlignTune 0.1.28 and AuditKIT 1.1.3. Exact dependency locks and protocols are in the source repository; historical environments are recorded in their original reports. Frozen final inference uses FP16 compute, NF4 double quantization and the pinned model revision. Successful loading alone does not reproduce the measured scoring protocol.