text stringlengths 3 8.33k | repo stringclasses 52
values | path stringlengths 6 141 | language stringclasses 35
values | sha stringlengths 64 64 | chunk_index int32 0 273 | n_tokens int32 1 896 |
|---|---|---|---|---|---|---|
# Phase 4: Synthetic Dataset Generation - Research
## RESEARCH COMPLETE
**Question answered:** What needs to be known to plan Phase 4 well?
## Scope Summary
Phase 4 creates the synthesis subsystem that consumes immutable Phase 3 corpus artifacts and emits canonical training datasets. The required outputs are conten... | adapter-lab | .planning/phases/04-synthetic-dataset-generation/04-RESEARCH.md | Markdown | fd132f08d50bfe2502d1be39dd8e532cd3ac225fb565370e7a013d6ffc4ab043 | 0 | 896 |
model refs, filter evidence, canary evidence, and `SynthManifest`.
- Strategy tests for deterministic Bonito-compatible, Magpie-style, Self-Instruct-compatible, human-seed, and preference-pair generation.
- Resolver tests for local refs, denied airgapped egress, restricted allowlists, connected allowlists, and manifest... | adapter-lab | .planning/phases/04-synthetic-dataset-generation/04-RESEARCH.md | Markdown | ba4843c00e550759c106baff519772fdd25438ca2b76d6e1cbc85bb8a05f23c2 | 1 | 550 |
---
phase: 04
slug: synthetic-dataset-generation
status: verified
nyquist_compliant: true
wave_0_complete: true
created: 2026-05-26
---
# Phase 04 - Validation Strategy
> Per-phase validation contract for feedback sampling during execution.
---
## Test Infrastructure
| Property | Value |
|----------|-------|
| **F... | adapter-lab | .planning/phases/04-synthetic-dataset-generation/04-VALIDATION.md | Markdown | 768968dadf9cb05028508254bd2652e858a09c5cbe865608a998e6686956b2f1 | 0 | 896 |
] Synthesis contract schemas are exported and included in schema drift tests.
- [x] Socket-blocking no-network fixture covers `adapterlab synthesize --json` in airgapped mode.
- [x] Privacy sentinel strings include PII-like values, canary-like values, and secret-like values.
---
## Manual-Only Verifications
Manual r... | adapter-lab | .planning/phases/04-synthetic-dataset-generation/04-VALIDATION.md | Markdown | ab29e492e13066dba5fcfb5b8fb54a1c2993766740704685afb14e3cb3f13673 | 1 | 206 |
---
phase: 04-synthetic-dataset-generation
status: passed
verified: 2026-05-27
requirements:
accounted_for:
- SYN-01
- SYN-02
- SYN-03
- SYN-04
- SYN-05
- SYN-06
- SYN-07
- SYN-08
- SYN-09
- SYN-10
- SYN-11
- SYN-12
human_verification: []
gaps: []
---
# Phase 04 Verifi... | adapter-lab | .planning/phases/04-synthetic-dataset-generation/04-VERIFICATION.md | Markdown | 854cea7c6a87a3dfa5abd88dfeac716a4c935572d7c3047131052daf397e7234 | 0 | 372 |
---
phase: 05-adapter-training
plan: 05-01
title: Base model registry, training config, local model guardrails, license metadata, and deterministic-mode controls
type: tdd
wave: 1
depends_on: []
files_modified:
- src/adapterlab/contracts/training.py
- src/adapterlab/contracts/__init__.py
- src/adapterlab/contract... | adapter-lab | .planning/phases/05-adapter-training/05-01-PLAN.md | Markdown | c8d96b9ffc52419bd40022795a48bf360ab2ed3a3b950f8be4c5315c8b189666 | 0 | 896 |
-training/05-RESEARCH.md`
- `.planning/phases/05-adapter-training/05-VALIDATION.md`
- `src/adapterlab/contracts/synthesis.py`
- `src/adapterlab/contracts/artifacts.py`
- `src/adapterlab/contracts/schema_export.py`
- `src/adapterlab/contracts/validation.py`
- `tests/contracts/test_syn... | adapter-lab | .planning/phases/05-adapter-training/05-01-PLAN.md | Markdown | 56b97f3a332a5cda1edd41eb1e10683a352892e610ba9c88db4fa8fb95fbc735 | 1 | 896 |
uv run pytest tests/contracts/test_training_models.py tests/contracts/test_schema_exports.py -q</verify>
</task>
<task id="05-01-03" type="green">
<name>Implement train config loader and safe template</name>
<read_first>
- `tests/training/test_config.py`
- `src/adapterlab/synthesis/config.py`
... | adapter-lab | .planning/phases/05-adapter-training/05-01-PLAN.md | Markdown | cca6d5d3339a74949e27d0d51742dbedd91a737c97f526655c692705db180d62 | 2 | 896 |
tests/contracts/test_schema_exports.py`
</read_first>
<action>Regenerate committed JSON Schemas so `schemas/train-config.schema.json` and `schemas/train-manifest.schema.json` exist. Run formatting and focused contract/config/base-model verification, and add a regression assertion that importing training contrac... | adapter-lab | .planning/phases/05-adapter-training/05-01-PLAN.md | Markdown | 02cd1fa1c043cc4e4eedc135a6a5e164a66823ca2b110fa7ba381e0b2f617e25 | 3 | 407 |
---
phase: 05-adapter-training
plan: 05-01
subsystem: training
tags: [contracts, config, base-models, schemas, deterministic-mode]
requires:
- phase: 04-synthetic-dataset-generation
provides: content-addressed SFT and preference dataset refs
provides:
- Strict TrainConfig and TrainManifest contracts
- BaseMod... | adapter-lab | .planning/phases/05-adapter-training/05-01-SUMMARY.md | Markdown | fb468900cfdc44d0004c4fe132440d79a249de95d4b4c7c39c777c91eea2c89b | 0 | 796 |
---
phase: 05-adapter-training
plan: 05-02
title: LoRA, QLoRA, and 8-bit LoRA trainer paths with PEFT adapter output
type: tdd
wave: 2
depends_on: [05-01]
files_modified:
- src/adapterlab/training/datasets.py
- src/adapterlab/training/methods.py
- src/adapterlab/training/backends/__init__.py
- src/adapterlab/tr... | adapter-lab | .planning/phases/05-adapter-training/05-02-PLAN.md | Markdown | 1d2b5a6a1752813124f229ac33f4a48b2f455f0bb33cb3d8cf966d2070dbdf12 | 0 | 896 |
bit method configs through a common registry, fixture backend adapter output layout, `run_training()` result shape, dry-run behavior, and `adapterlab train --config PATH --json` machine-readable output.</action>
<acceptance_criteria>
- `tests/training/test_datasets.py` asserts SFT records load from a local ar... | adapter-lab | .planning/phases/05-adapter-training/05-02-PLAN.md | Markdown | a1eea4630fdfa322a8215c367e560c98a1bdc5e412ff59338a8b1c6a17520824 | 1 | 896 |
does not contain SFT instructions or responses.
- `uv run pytest tests/training/test_lora_backends.py -q` exits 0.
</acceptance_criteria>
<verify>uv run pytest tests/training/test_lora_backends.py -q</verify>
</task>
<task id="05-02-04" type="green">
<name>Add lazy Hugging Face LoRA backend shell... | adapter-lab | .planning/phases/05-adapter-training/05-02-PLAN.md | Markdown | 9d745bf3b511b66d07009efe8cdff5fcf5366743043a58aeecfbc295c53b9bdd | 2 | 896 |
adapterlab/cli.py tests/training tests/cli/test_training_cli.py`
</verification>
<success_criteria>
- LoRA, QLoRA, and 8-bit LoRA method configs run through one training abstraction.
- Default fixture training writes PEFT-compatible adapter layout refs.
- Real Hugging Face backend hooks are local-only and lazy-importe... | adapter-lab | .planning/phases/05-adapter-training/05-02-PLAN.md | Markdown | d0bff6019b612af859fa836b77dcb465785b33ec9f13e83647e06e249b499f44 | 3 | 87 |
---
phase: 05-adapter-training
plan: 05-02
subsystem: training
tags: [lora, qlora, peft, fixture-backend, cli]
requires:
- phase: 05-adapter-training
provides: training contracts and local base model resolution
provides:
- Dataset loaders for SFT and preference JSONL refs
- Training method registry for LoRA-f... | adapter-lab | .planning/phases/05-adapter-training/05-02-SUMMARY.md | Markdown | a9ff8ebc7376fb8c5800b0dd89eb1162fb390ceb99666f9602b83329d5c30606 | 0 | 836 |
---
phase: 05-adapter-training
plan: 05-03
title: DPO and DP-LoRA method adapters with reference model handling and RDP accounting
type: tdd
wave: 2
depends_on: [05-01, 05-02]
files_modified:
- src/adapterlab/contracts/training.py
- src/adapterlab/training/datasets.py
- src/adapterlab/training/methods.py
- src/... | adapter-lab | .planning/phases/05-adapter-training/05-03-PLAN.md | Markdown | 7c2f813b467e442e7e03fe9487d45cf6132dff8687aebd01a7955f9a7aa9e001 | 0 | 896 |
`accountant`.
- `tests/training/test_dpo_backend.py` asserts projected DPO rows contain `prompt`, `chosen`, `rejected`, and internal ID only.
- `tests/training/test_dp_lora.py` asserts DP evidence has accounting fields and no raw sample values.
- `tests/privacy/test_training_outputs.py` includes prefe... | adapter-lab | .planning/phases/05-adapter-training/05-03-PLAN.md | Markdown | dfe68b77e115d798a78f311271f0b642a37bb85a2dbecadfd4b01fc4c604843d | 1 | 896 |
training/05-RESEARCH.md`
- `src/adapterlab/training/backends/hf_lora.py`
- `src/adapterlab/training/backends/dp.py`
- `src/adapterlab/training/base_model.py`
- `pyproject.toml`
</read_first>
<action>Create `src/adapterlab/training/backends/hf_dpo.py` with lazy imports for `trl`, `transfo... | adapter-lab | .planning/phases/05-adapter-training/05-03-PLAN.md | Markdown | eeaa64549f430c15a2a6a9302ca85870ca9468aef674ee477ab6c147810c3203 | 2 | 761 |
---
phase: 05-adapter-training
plan: 05-03
subsystem: training
tags: [dpo, dp-lora, privacy, rdp, preference-training]
requires:
- phase: 05-adapter-training
provides: method registry and fixture training pipeline
provides:
- DPO preference projection and reference model evidence
- DP-LoRA aggregate RDP-style... | adapter-lab | .planning/phases/05-adapter-training/05-03-SUMMARY.md | Markdown | 61034271cf4b2b2930f3d5f0eb2a0e8741c7cf31bf5281dd80f70352f5bc7ac7 | 0 | 725 |
---
phase: 05-adapter-training
plan: 05-04
title: FSDP/FSDP2 or configured backend integration, hardware capture, and training manifest completion
type: tdd
wave: 3
depends_on: [05-02, 05-03]
files_modified:
- src/adapterlab/training/hardware.py
- src/adapterlab/training/manifest.py
- src/adapterlab/training/pipe... | adapter-lab | .planning/phases/05-adapter-training/05-04-PLAN.md | Markdown | 4f4c99f789ceb7fea2893c05d174d7990ae73ba9154ab26f7f1bc88809f71fae | 0 | 896 |
training/test_manifest.py` asserts `TrainManifest` includes inputs, configs, base model, method, backend, hardware, libraries, metrics, outputs, seeds, determinism, timing, egress decisions, and plugin/backend evidence.
- `tests/training/test_pipeline.py` asserts `manifest_ref` resolves to JSON with stage_id `tra... | adapter-lab | .planning/phases/05-adapter-training/05-04-PLAN.md | Markdown | eb5afb7e90a1de371cb8ac0b2b3310ac7560a093eaf2661ee83deb7085afe423 | 1 | 896 |
05-04-04" type="green">
<name>Add local metrics and experiment evidence for training</name>
<read_first>
- `tests/observability/test_experiments.py`
- `tests/observability/test_metrics.py`
- `src/adapterlab/observability/metrics.py`
- `src/adapterlab/observability/experiments.py`
-... | adapter-lab | .planning/phases/05-adapter-training/05-04-PLAN.md | Markdown | 85f6e22b8786ae1fa46baf58686b0e2d80e84b0323ecf6c4fc1cd1e40763eeb0 | 2 | 782 |
---
phase: 05-adapter-training
plan: 05-04
subsystem: training
tags: [manifest, hardware, metrics, experiments, fsdp-evidence]
requires:
- phase: 05-adapter-training
provides: training method evidence for LoRA, DPO, and DP-LoRA
provides:
- Hardware and optional library capture
- Complete train.manifest.json a... | adapter-lab | .planning/phases/05-adapter-training/05-04-SUMMARY.md | Markdown | 18b41d03b09664b37f80e4a457af37028e531571a20d48cc765ef96bd42b91d6 | 0 | 715 |
---
phase: 05-adapter-training
plan: 05-05
title: Tokenizer and merge provenance, PEFT load round-trip, tiny-model tests, and training privacy checks
type: tdd
wave: 4
depends_on: [05-04]
files_modified:
- src/adapterlab/contracts/training.py
- src/adapterlab/training/verify.py
- src/adapterlab/training/manifest.... | adapter-lab | .planning/phases/05-adapter-training/05-05-PLAN.md | Markdown | 577fcc80bf62bfb0be749161c1b756f5a8f5c1218c11fb1bc779e7f8ed29db07 | 0 | 896 |
logs.</action>
<acceptance_criteria>
- `tests/training/test_peft_roundtrip.py` asserts fixture adapter dirs contain required PEFT filenames.
- `tests/training/test_peft_roundtrip.py` defines an optional marked real PEFT tiny-model test that uses local/generated fixtures only.
- `tests/training/tes... | adapter-lab | .planning/phases/05-adapter-training/05-05-PLAN.md | Markdown | b304ef5b194db608bd0f5c6222790607fdc92a500ed906a2764abe41d0ac0688 | 1 | 896 |
-05-04" type="green">
<name>Add train no-network and privacy regression coverage</name>
<read_first>
- `tests/offline/test_no_network.py`
- `tests/privacy/test_training_outputs.py`
- `tests/privacy/test_enterprise_outputs.py`
- `src/adapterlab/security/redaction.py`
- `src/adapterl... | adapter-lab | .planning/phases/05-adapter-training/05-05-PLAN.md | Markdown | fd41ce750c998864e80c29ee6e06f5e36a615e6d01660871f958157f333d697f | 2 | 896 |
Adapter layout verification is automated.
- Optional real PEFT tiny-model round-trip is documented and locally runnable.
- Tokenizer extension and merge provenance are manifest-recorded.
- Training has no-network and privacy-regression coverage.
- Phase 5 is ready for `$gsd-execute-phase 5`.
</success_criteria>
| adapter-lab | .planning/phases/05-adapter-training/05-05-PLAN.md | Markdown | 85f46ecc284c978ebb33e95b5beb7d54ce2a0b91d7df1991d5dbbba5eae149fd | 3 | 65 |
---
phase: 05-adapter-training
plan: 05-05
subsystem: training
tags: [peft, adapter-verification, tokenizer, merge, privacy, no-network]
requires:
- phase: 05-adapter-training
provides: complete train manifest and fixture training outputs
provides:
- Adapter layout verification
- Tokenizer extension and adapt... | adapter-lab | .planning/phases/05-adapter-training/05-05-SUMMARY.md | Markdown | a351bba910f551080096f3a1bcc22775950afff5e0351ec6d1aece490739fabd | 0 | 836 |
# Phase 5: Adapter Training - Context
**Gathered:** 2026-05-27
**Status:** Ready for planning
<domain>
## Phase Boundary
Phase 5 delivers the training subsystem that consumes Phase 4 `sft.jsonl`,
optional `pref.jsonl`, and `synth.manifest.json` artifact refs, resolves local
open-weight base models, runs configured P... | adapter-lab | .planning/phases/05-adapter-training/05-CONTEXT.md | Markdown | dfb1579d7292c6f55ba95cbdfded84ca27ba0f4e681fcd067cc34131bc042144 | 0 | 896 |
/metrics refs, wall-clock timing, output refs, plugin/backend evidence,
and egress decisions.
- **D-14:** Deterministic mode is opt-in. When enabled, record deterministic
flags, seeds, backend settings, and incompatible kernels or known sources of
nondeterminism. Do not claim bit-for-bit reproducibility unless th... | adapter-lab | .planning/phases/05-adapter-training/05-CONTEXT.md | Markdown | e15b427e93c17c4d0e4bb1607f8e48f6996e23e753834491471997e0d9c6440e | 1 | 896 |
dataset-generation/04-VERIFICATION.md` -
Verification evidence showing Phase 4 outputs are ready for Phase 5.
### Existing Code
- `pyproject.toml` - Current package dependencies are intentionally light:
Pydantic, Typer, PyYAML, and PyArrow only; heavy ML dependencies are not in
the base install.
- `src/adapterla... | adapter-lab | .planning/phases/05-adapter-training/05-CONTEXT.md | Markdown | fbf4139aa1c7b7763b136f1168ce84288a7c2dad1346c2a023be7e3e947b52ec | 2 | 896 |
default local training fixture.
</code_context>
<specifics>
## Specific Ideas
- Recommended defaults were selected for all five gray areas after the user
explicitly instructed: "Always choose recommended and continue."
- Keep default repository verification practical on machines without CUDA or
ML packages, whil... | adapter-lab | .planning/phases/05-adapter-training/05-CONTEXT.md | Markdown | aaf90bbb5c0e3234a4cc216a14818d77e094a84831993cedd9ee073137c369c9 | 3 | 158 |
# Phase 5: Adapter Training - Discussion Log
> **Audit trail only.** Do not use as input to planning, research, or execution agents.
> Decisions are captured in CONTEXT.md — this log preserves the alternatives considered.
**Date:** 2026-05-27
**Phase:** 5-Adapter Training
**Areas discussed:** Base model guardrails, T... | adapter-lab | .planning/phases/05-adapter-training/05-DISCUSSION-LOG.md | Markdown | 2b64cf7c34b4f6b6320ce58103d0cb4cdd88e333722cffe56f8f34a26ff4d1f4 | 0 | 896 |
---------|
| Strict PEFT-compatible output contract | Produce `adapter_model.safetensors`, `adapter_config.json`, tokenizer artifacts only when needed, and adapter load round-trip verification. | ✓ |
| Manifest-only output first | Emit manifest and placeholder output refs before enforcing PEFT layout. | |
| Backend-spe... | adapter-lab | .planning/phases/05-adapter-training/05-DISCUSSION-LOG.md | Markdown | 819b86ded5bbe5ed780da0903736a3d5e6896a050335dd52d594b00a36ec9a7c | 1 | 211 |
# Phase 05: Adapter Training - Pattern Map
**Mapped:** 2026-05-27
**Files analyzed:** 33 new/modified files
**Analogs found:** 33 / 33
## File Classification
| New/Modified File | Role | Data Flow | Closest Analog | Match Quality |
|-------------------|------|-----------|----------------|---------------|
| `src/adap... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | 5c098649f240c5b86d63d655163d747c61879dfb5788e56f653e7a307ba864ec | 0 | 896 |
| config | file-I/O | `schemas/synth-config.schema.json` | exact |
| `schemas/train-manifest.schema.json` | config | file-I/O | `schemas/synth-manifest.schema.json` | exact |
| `pyproject.toml` | config | transform | `pyproject.toml` dependency/test config | role-match |
| `tests/contracts/test_training_models.py` | te... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | 50e7c5ca2d623710ec69e9117d18f0983170d11b71b8359a41072992fb586ffa | 1 | 896 |
None
quality: list[QualityFilterResult] = Field(default_factory=list)
metadata: dict[str, Any] = Field(default_factory=dict)
```
```python
class PreferenceRecord(StrictContract):
"""Canonical preference-pair example for DPO-style training."""
schema_version: str = "preference-record/v1"
example_id... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | 0abea87ec15d2bfb9c03d7fa06f1d829acb5ca67b5e80917d1f90d617c3ba21d | 2 | 896 |
* `src/adapterlab/synthesis/pipeline.py`
**SDK-first imports pattern** (lines 5-21):
```python
from dataclasses import asdict, dataclass
from pathlib import Path
from adapterlab.artifacts import LocalArtifactStore
from adapterlab.contracts import ArtifactRef
from adapterlab.contracts.synthesis import SFTRecord
from a... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | 618d6748ee36ad19d0c540205f79b2c1cfacffa1079dc7279bdb7e5a92d33fdf | 3 | 896 |
= None,
) -> SynthManifest:
"""Build a complete synthesis manifest for stage handoff."""
inputs = {"corpus": config.corpus}
if config.ingest_manifest is not None:
inputs["ingest_manifest"] = config.ingest_manifest
configs = {"synth": config_ref} if config_ref is not None else {}
egress_deci... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | 23878c278b2fed1eebb331dd506113ca1fd6ae802c4542173d2066ccaf02c48a | 4 | 896 |
).hexdigest()
if actual != ref.digest:
raise ContractValidationError(
f"Artifact digest mismatch: expected {ref.digest}, got {actual}",
field_path="digest",
)
```
**Apply to:** base model guardrail implementation. Verify local path or `ArtifactRef` bytes before any backend s... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | c0707002e0f21e97bad591ad74fbe697a6663b897c7b8a111a3eb00bae877300 | 5 | 896 |
`, or package import paths. Missing optional dependencies should raise stable `ConfigError` or capability/preflight errors.
---
### `src/adapterlab/training/hardware.py` and `src/adapterlab/training/verify.py` (utility/service)
**Analog:** `src/adapterlab/doctor/checks.py`, `src/adapterlab/artifacts/store.py`, and e... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | 4481e49264e70306e360e60d58c321085836539cd4d9310f30f654d2090f0c02 | 6 | 896 |
config, request-response)
**Analog:** `src/adapterlab/synthesis/__init__.py`
**SDK export pattern** (lines 1-6):
```python
"""Synthetic dataset generation SDK."""
from adapterlab.synthesis.config import load_synth_config
from adapterlab.synthesis.pipeline import SynthesisResult, run_synthesis
__all__ = ["SynthesisR... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | cfda6543641548304d39c911ca0cd3d0e30114f65208945b911547ec97f27149 | 7 | 896 |
` imports, `train-config`/`train-manifest` validation kinds, and committed `schemas/train-config.schema.json` plus `schemas/train-manifest.schema.json`.
---
### `src/adapterlab/templates/train.yaml` (config, file-I/O)
**Analog:** `src/adapterlab/templates/synth.yaml`
**Safe local template pattern** (lines 1-29):
``... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | 773b5556bfb84bbce60914969d1690f8b723d3d0fd73dbc6a9032c92b0c3b4c2 | 8 | 896 |
" in payload["configs"]
assert payload["strategies"] == {"bonito_grounded": 1}
assert "generator" in payload["models"]
assert "sft" in payload["outputs"]
assert payload["seeds"]["run"] == 1234
assert payload["license_metadata"]["generator"] == "apache-2.0"
assert payload["quality"]["filter_count... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | b9b43f1c82df30f597e552b5c4f5dba91327acfbb53e7804197332340f62b0a7 | 9 | 896 |
tenant-config.schema.json",
"synth-config.schema.json",
"synth-manifest.schema.json",
}
assert expected.issubset({path.name for path in SCHEMA_DIR.glob("*.schema.json")})
def test_fnd_09_schema_exports_match_committed_files(tmp_path) -> None:
export_schemas(tmp_path)
for name, schema... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | 237ccce7698a3495744396f3e7c68a81162cb31cc796af3228ad61d2403bcc22 | 10 | 896 |
Public SDK Exports
**Source:** `src/adapterlab/synthesis/__init__.py` and `src/adapterlab/contracts/__init__.py`
**Apply to:** `adapterlab.training` and public training contracts.
```python
from adapterlab.synthesis.config import load_synth_config
from adapterlab.synthesis.pipeline import SynthesisResult, run_synthesi... | adapter-lab | .planning/phases/05-adapter-training/05-PATTERNS.md | Markdown | 1e76e9fc5086fdfee0e86811f20c3a114c8d58d3a4e3f005736df06f4a269fca | 11 | 594 |
# Phase 05: Adapter Training - Research
**Researched:** 2026-05-27
**Domain:** Offline PEFT adapter training, provenance manifests, local model guardrails
**Confidence:** HIGH for architecture and local guardrails, MEDIUM-HIGH for ML backend APIs
<user_constraints>
## User Constraints (from CONTEXT.md)
### Locked De... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 3675c993c899fbb33a8cb31b747bad4bfcbaaa4ccb2bf05216bfe33f1dea0486 | 0 | 896 |
epochs, and resulting RDP or epsilon summary.
- **D-16:** Privacy evidence remains aggregate and reference-based. Logs, CLI
JSON, manifests, metrics, and experiment records must not expose raw SFT or
preference text, PII values, canary strings, signing secrets, HSM PINs, or
private keys.
### Adapter Output Contr... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 894dbd159cf986f08a673e5485ba2939658b864a82ace32cc5b1e7ee31955702 | 1 | 896 |
Description | Research Support |
|----|-------------|------------------|
| TRN-01 | Operator can train a PEFT-compatible LoRA adapter from local open-weight base model files. | PEFT `LoraConfig`, `get_peft_model`, and `save_pretrained` support the adapter output contract. [CITED: https://huggingface.co/docs/peft/main/e... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 089749930d48d504d967fdcc3b68911d9d4b5e927e61462ab4b3e93b758b74ad | 2 | 896 |
.py] |
| TRN-10 | A trained adapter can be loaded round-trip through standard PEFT APIs in an automated verification test. | PEFT `PeftModel.from_pretrained` loads a saved adapter directory into a base model; the base model must be supplied locally. [CITED: https://huggingface.co/docs/peft/main/en/package_reference/pef... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 60815ea5af9f6d6c8fbb17b0905ad76b69df0f089cbf905b3114b67d7d52e455 | 3 | 896 |
phases/05-adapter-training/05-CONTEXT.md]
## Architectural Responsibility Map
| Capability | Primary Tier | Secondary Tier | Rationale |
|------------|--------------|----------------|-----------|
| `train.yaml` parsing and validation | API / Backend | CLI | SDK contracts own validation; Typer should remain a thin wra... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | a5b2ea702dafd8b4b07fcdaf8247ea62cc6e85181f3922e3ed392b0ffde730d4 | 4 | 896 |
------|---------|---------|-------------|
| `torch` | Latest PyPI 2.12.0, installed audit 2.12.0 | Tensor/training runtime, deterministic controls, FSDP/FSDP2 | Required for real training backends and PEFT load round-trip. [VERIFIED: PyPI + slopcheck] [CITED: https://docs.pytorch.org/docs/2.12/notes/randomness.html] |
... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 721f3130e711302fc083ec27cce6d47dde26636d6886547faa3f6ab657027dc0 | 5 | 896 |
: https://huggingface.co/docs/trl/en/dpo_trainer] |
| `bitsandbytes` | GPTQ/AWQ/torchao quantization | Phase scope specifically requires QLoRA NF4 and 8-bit LoRA; bitsandbytes is the documented Transformers path for those features. [CITED: https://huggingface.co/docs/transformers/main/quantization/bitsandbytes] |
**In... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 3aa959fd406d0e5eca3a4d0cfe56972acad2993178d6f8fa83bf0d26423fbcc1 | 6 | 896 |
0.49.2 uploaded 2026-02-16 | Not exposed by PyPI JSON; not checked | PyPI JSON source URL unavailable; official docs link to GitHub project | OK | Approved with hardware preflight gating. [VERIFIED: PyPI + slopcheck] [CITED: https://huggingface.co/docs/transformers/main/quantization/bitsandbytes] |
| `opacus` | PyPI | ... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 1f5a7ecc99de99f5fd0fb4010eca1752924190d6283b5d452f878501bba20ea6 | 7 | 896 |
.kind]
evidence = method.run(TrainingContext(config=config, base_model=base_model, datasets=datasets))
manifest = build_manifest(config=config, base_model=base_model, evidence=evidence)
manifest_ref = write_manifest(manifest, store)
return TrainingResult(manifest_ref=manifest_ref.model_dump(mode="json")... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 605201078da17923f2bc3c6453d28b7dc14b5e4191014648878cee88bdabbe61 | 8 | 896 |
across releases/platforms. [CITED: https://docs.pytorch.org/docs/2.12/notes/randomness.html]
- **Storing optimizer state bytes directly in manifests:** Store a hash/ref; manifests should stay durable, bounded, and privacy-safe. [CITED: AGENTS.md]
- **Logging SFT/preference record text:** Phase 5 evidence must remain ag... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 9200c91682055fb38adca4e02b9505da8fbc92340540bdab1b7f24f5c9dbf283 | 9 | 896 |
store.py]
**Warning signs:** `train.yaml` has only `model_name_or_path` or a Hub repo ID; no `sha256:` fields exist. [ASSUMED]
### Pitfall 2: QLoRA/8-bit Paths Passing on CPU Fixture but Failing on Real CUDA
**What goes wrong:** Tests pass with fixture output, but real `bitsandbytes` backend fails because platform/har... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 913daa5111a728a5430406691fc65f9079a2da46524e45496d36b2cd856dfdec | 10 | 896 |
record.prompt,
"chosen": record.chosen,
"rejected": record.rejected,
}
```
### DP Accounting Evidence
```python
# Source: Opacus PrivacyEngine and RDPAccountant docs
privacy_engine = PrivacyEngine(accountant="rdp", secure_mode=config.secure_mode)
model, optimizer, dataloader = privacy_engine.make_p... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | fdc8b3bd2c3897c7196b94b280946cc3f69b8740eb5c7291314e3cf0d14083b9 | 11 | 896 |
omit quantization preflight and optional markers. |
| A4 | Raw Phase 4 preference dictionaries passed to backend callbacks could leak unused fields. | Common Pitfalls | Planner may need explicit dataset projection tasks. |
| A5 | DP configs with only `enabled: true` are insufficient. | Common Pitfalls | Planner may nee... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 014ee624da3e63d1efa929addef84355cb291530773ba6d88dbd4a77e540584b | 12 | 896 |
importlib.metadata] |
| `bitsandbytes` | QLoRA / 8-bit LoRA | yes in agent shell after package audit | 0.49.2 | Capability-gate and skip real quantization on unsupported hardware. [VERIFIED: importlib.metadata] |
| `opacus` | DP-LoRA accounting | yes in agent shell after package audit | 1.6.0 | Fixture accounting evide... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 81302023a7041a81a7dfaa06dcf354d1337b62ffd449351cf37501a124d46138 | 13 | 896 |
| `uv run pytest tests/training/test_determinism.py -q` | no, Wave 0 |
| TRN-13 | Tokenizer/merge provenance | unit + optional integration | `uv run pytest tests/training/test_tokenizer_merge.py -q` | no, Wave 0 |
| OPS-10 | Tiny-model training and adapter load test | optional integration | `uv run pytest tests/trainin... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | 834c5c3ccae85319f591b6810e2019abbb180b4b58fc9323915307fc1ae63477 | 14 | 896 |
] |
### Known Threat Patterns for Offline Training Stack
| Pattern | STRIDE | Standard Mitigation |
|---------|--------|---------------------|
| Training-time model fetch in airgapped mode | Information Disclosure / Policy Bypass | Pre-resolve local base model, verify digests, set offline/local flags, extend socket-b... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | aa2c890c9a8e71a2f5cbfb45b2a6a2f72ce8321908dbb595340f7e49232fbabe | 15 | 896 |
main/en/installation#offline-mode] [CITED: https://huggingface.co/docs/transformers/main/quantization/bitsandbytes]
- Hugging Face TRL DPO docs - DPO dataset format, metrics, PEFT integration. [CITED: https://huggingface.co/docs/trl/en/dpo_trainer]
- Hugging Face Accelerate FSDP docs - single-node/FSDP config, state sa... | adapter-lab | .planning/phases/05-adapter-training/05-RESEARCH.md | Markdown | e0a707e956fcb82610071c1dea15ce1768c909659d929d0c09fe39b8cede2004 | 16 | 504 |
---
phase: 05
slug: adapter-training
status: verified
threats_open: 0
asvs_level: 3
created: 2026-05-28T20:05:00Z
updated: 2026-05-28T20:05:00Z
---
# Phase 05 — Security
Per-phase security contract for Phase 05: Adapter Training.
## Trust Boundaries
| Boundary | Description | Data Crossing |
|----------|-----------... | adapter-lab | .planning/phases/05-adapter-training/05-SECURITY.md | Markdown | 06adf9c2a218ce944e826c13e0fbd8a6e491ab149a85b89cddfdb700787bc87c | 0 | 896 |
and real backend capability checks fail with stable `ConfigError` when dependencies are absent. | closed |
| T-05-HEAVY-IMPORT | Denial of Service | Optional LoRA backend | mitigate | Hugging Face LoRA imports remain lazy in `hf_lora.py`, while fixture backend remains default for offline tests. | closed |
| T-05-DPO-SH... | adapter-lab | .planning/phases/05-adapter-training/05-SECURITY.md | Markdown | 443d21a71f34c1c29f490589cc057490564863e7915b4818a84e9dc689fc9146 | 1 | 896 |
, `tests/privacy/test_training_outputs.py`, `tests/offline/test_no_network.py`, `tests/cli/test_training_cli.py`, and `tests/contracts/test_training_models.py`.
- Phase 05 summaries record focused and full-suite verification, including `uv run pytest -q`, `uv run ruff check .`, and representative `adapterlab train --co... | adapter-lab | .planning/phases/05-adapter-training/05-SECURITY.md | Markdown | 39d1f642cec1c6aa2e2c9a7ffdf9b719bfc67ac5fffd6058032702107f9f8e83 | 2 | 159 |
---
status: complete
phase: 05-adapter-training
source:
- 05-01-SUMMARY.md
- 05-02-SUMMARY.md
- 05-03-SUMMARY.md
- 05-04-SUMMARY.md
- 05-05-SUMMARY.md
started: 2026-05-28T19:30:00Z
updated: 2026-05-28T19:55:00Z
---
## Current Test
[testing complete]
## Tests
### 1. Train CLI Produces Privacy-Safe Adapter ... | adapter-lab | .planning/phases/05-adapter-training/05-UAT.md | Markdown | 2aef31a22c09017c2c1f420760e3980d4021268e26c6cc6df54b4c2eaa2c6e04 | 0 | 611 |
---
phase: 05
slug: adapter-training
status: verified
nyquist_compliant: true
wave_0_complete: true
created: 2026-05-27
last_audited: 2026-05-28
---
# Phase 05 - Validation Strategy
> Per-phase validation contract for feedback sampling during execution.
---
## Test Infrastructure
| Property | Value |
|----------|-... | adapter-lab | .planning/phases/05-adapter-training/05-VALIDATION.md | Markdown | 88c201b957fb6f252e86149ef39c87ce0398919abdd51dd02e95d01c3c39302e | 0 | 896 |
-DPO-SHAPE-DRIFT, T-05-MISSING-REF-MODEL, T-05-UNSUPPORTED-DP-CLAIM, T-05-PREF-LEAK | DPO requires preference refs and reference model metadata; DP-LoRA records aggregate DP accounting only | unit/contract/privacy | `uv run pytest tests/contracts/test_training_models.py tests/training/test_dpo_backend.py tests/training... | adapter-lab | .planning/phases/05-adapter-training/05-VALIDATION.md | Markdown | b0e6085de8d36a40d487392892683586fa0180c74d0d011bf359b4fd812ac1de | 1 | 896 |
-|
| Gaps found | 0 |
| Resolved | 0 |
| Escalated | 0 |
| Automated rows green | 6 |
| Manual-only rows retained | 2 |
Focused evidence:
- `uv run pytest tests/training tests/contracts/test_training_models.py tests/contracts/test_schema_exports.py tests/offline/test_no_network.py tests/privacy/test_training_outputs.... | adapter-lab | .planning/phases/05-adapter-training/05-VALIDATION.md | Markdown | 279b6ff1b79c4ca74000ade7771dd4c073a329f58cd90207f022ddc1c3a970f5 | 2 | 94 |
---
phase: 05-adapter-training
status: passed
created: 2026-05-29
requirements:
accounted_for:
- TRN-01
- TRN-02
- TRN-03
- TRN-04
- TRN-05
- TRN-06
- TRN-07
- TRN-08
- TRN-09
- TRN-10
- TRN-11
- TRN-12
- TRN-13
- OPS-10
human_verification: []
gaps: []
---
# Ph... | adapter-lab | .planning/phases/05-adapter-training/05-VERIFICATION.md | Markdown | c37327387de2db1bc153d33a72ef2e76d1242c16b955e6475eb09f39be2a8180 | 0 | 684 |
---
phase: 06-evaluation-reports
plan: 06-01
title: Evaluator protocol, lm-evaluation-harness integration, dataset refs, grader refs, and local resource policy
type: tdd
wave: 1
depends_on: []
files_modified:
- src/adapterlab/contracts/evaluation.py
- src/adapterlab/contracts/__init__.py
- src/adapterlab/contract... | adapter-lab | .planning/phases/06-evaluation-reports/06-01-PLAN.md | Markdown | fe53d744cb48353aaf5f7749c5123bc34ab268f273659207535ef06bd952ed52 | 0 | 896 |
task id="06-01-01" type="red">
<name>Add failing evaluation contract, schema, and config tests</name>
<read_first>
- `.planning/phases/06-evaluation-reports/06-CONTEXT.md`
- `.planning/phases/06-evaluation-reports/06-RESEARCH.md`
- `.planning/phases/06-evaluation-reports/06-VALIDATION.md`
... | adapter-lab | .planning/phases/06-evaluation-reports/06-01-PLAN.md | Markdown | 65d2a90440da45f2f09cffd2fde3754bb87105aa9b45d64c4a1bfbe3e0f6a1db | 1 | 896 |
pytest tests/contracts/test_evaluation_models.py tests/contracts/test_schema_exports.py -q</verify>
</task>
<task id="06-01-03" type="green">
<name>Implement config loader, safe template, and fixture eval workspace</name>
<read_first>
- `tests/evaluation/test_config.py`
- `src/adapterlab/traini... | adapter-lab | .planning/phases/06-evaluation-reports/06-01-PLAN.md | Markdown | c19caa4143b8d8123e5aa423224a77e75fe8935eaf639a3c8ff49ba565fa907f | 2 | 896 |
and stable capability failure when the optional package is missing. The default test lane should validate argument construction and missing-extra behavior without importing or executing real `lm_eval`.</action>
<acceptance_criteria>
- Missing optional `lm_eval` raises a stable `ConfigError` or capability erro... | adapter-lab | .planning/phases/06-evaluation-reports/06-01-PLAN.md | Markdown | dd99f04955ce4a82f2c2a0143cb6c49677a5d9d3f114e8becb56e1a8972ea58e | 3 | 667 |
# 06-01 Summary - Evaluator Protocol and Local Resource Policy
**Status:** Complete
**Date:** 2026-05-28
## Delivered
- Added strict evaluation contracts for `EvalConfig`, suite/resource refs, metric/result records, report refs, human-review summaries, and `EvalManifest`.
- Added schema exports for `eval-config.sche... | adapter-lab | .planning/phases/06-evaluation-reports/06-01-SUMMARY.md | Markdown | 0ba5e171f16b7feee92c9d2640bdd7fb8f8cb0ff92a9ce482198e961357f6385 | 0 | 204 |
---
phase: 06-evaluation-reports
plan: 06-02
title: Multilingual, instruction-following, BFCL/tau-bench, tool-inventory, and safety suite plugins
type: tdd
wave: 2
depends_on: [06-01]
files_modified:
- src/adapterlab/contracts/evaluation.py
- src/adapterlab/evaluation/suites.py
- src/adapterlab/evaluation/builtin... | adapter-lab | .planning/phases/06-evaluation-reports/06-02-PLAN.md | Markdown | 4a1dfe0ef6788193fbbd76d8a6a458011d391727a69b0d3711190d8bacee0e0a | 0 | 896 |
, IFEval-compatible instruction checks, BFCL-compatible function call records, tau-bench-compatible agentic task records, tool inventory loading, and safety/refusal panel aggregate metrics.</action>
<acceptance_criteria>
- Tests assert every suite returns `EvaluatorResult` with stable category, suite ID, metr... | adapter-lab | .planning/phases/06-evaluation-reports/06-02-PLAN.md | Markdown | 9eb6d84aa3a4639fb7c753a1e404a20ed9ac3aa143efa278fdbfe4e8c9f2a608 | 1 | 896 |
/adapterlab/evaluation/resources.py`
- `src/adapterlab/security/redaction.py`
</read_first>
<action>Create `src/adapterlab/evaluation/safety.py` with XSTest-compatible, HarmBench-compatible, StrongREJECT-compatible, and OR-Bench-compatible fixture panel evaluators. They should read local sanitized fixture... | adapter-lab | .planning/phases/06-evaluation-reports/06-02-PLAN.md | Markdown | ead6741d6d713c2f9973a75c438a7aa7fa55a8d7b4aebf785ad6a664487379d9 | 2 | 701 |
# 06-02 Summary - Suite Families and Tool Inventory
**Status:** Complete
**Date:** 2026-05-28
## Delivered
- Added deterministic fixture suite execution for capability, multilingual, function-calling, agentic-compatible, safety/refusal, custom, and human-review categories.
- Preserved configured language subsets as ... | adapter-lab | .planning/phases/06-evaluation-reports/06-02-SUMMARY.md | Markdown | f25523b05864d3ef202c0dd53b59e0e912a6b79594418c2ee2a405053f76e26a | 0 | 161 |
---
phase: 06-evaluation-reports
plan: 06-03
title: Sanitized report bundle, base-vs-adapter comparison outputs, and evaluation manifest
type: tdd
wave: 3
depends_on: [06-01, 06-02]
files_modified:
- src/adapterlab/contracts/evaluation.py
- src/adapterlab/evaluation/pipeline.py
- src/adapterlab/evaluation/manifes... | adapter-lab | .planning/phases/06-evaluation-reports/06-03-PLAN.md | Markdown | b89991e0bc3d7e35babece10c71f838adf6e2a7530675f7f35c1d349c0567820 | 0 | 896 |
`src/adapterlab/cli.py`
- `tests/cli/test_training_cli.py`
- `tests/privacy/test_training_outputs.py`
</read_first>
<action>Create failing tests for `run_evaluation()`, static report generation, base-vs-adapter comparison deltas, `EvalManifest` content, and `adapterlab evaluate --config ... --json` ... | adapter-lab | .planning/phases/06-evaluation-reports/06-03-PLAN.md | Markdown | fef633936e699370804684f381a724461480697c9dde2c701cb6c57067f59f79 | 1 | 896 |
metric records, comparison deltas, caveats, and evidence refs.
- HTML report is static and contains escaped dynamic values.
- Reports redact secret-like values and omit raw prompt/response/training text by default.
- Report writer returns `ArtifactRef`s for both JSON and HTML artifacts.
- `uv ru... | adapter-lab | .planning/phases/06-evaluation-reports/06-03-PLAN.md | Markdown | f1c8578822638e67199fb97c3906ffd949829aac9e2a7409a07abfb4aa1cf3c6 | 2 | 896 |
/report/manifest/CLI/privacy/offline checks and lint over evaluation modules.</action>
<acceptance_criteria>
- `uv run pytest tests/evaluation/test_pipeline.py tests/evaluation/test_reports.py tests/evaluation/test_manifest.py tests/cli/test_evaluation_cli.py tests/contracts/test_schema_exports.py tests/offli... | adapter-lab | .planning/phases/06-evaluation-reports/06-03-PLAN.md | Markdown | ab5a3c1ac12f816cccccfb2666e2e4f9a66dc5143f9c89e447ee6b408c8a658c | 3 | 471 |
# 06-03 Summary - Reports, Comparison Outputs, and Eval Manifest
**Status:** Complete
**Date:** 2026-05-28
## Delivered
- Added `run_evaluation()` pipeline with dry-run support, suite execution, metric refs, static JSON/HTML report refs, base-vs-adapter comparison deltas, human-review evidence hooks, and `eval.manif... | adapter-lab | .planning/phases/06-evaluation-reports/06-03-SUMMARY.md | Markdown | cf90cf0c74170bf113fdb73080ac3aa0b8f77ed71ba34df1d02c433e48a6515a | 0 | 218 |
---
phase: 06-evaluation-reports
plan: 06-04
title: Tenant evaluator plugin support, optional local human-review workflow, and evaluation privacy tests
type: tdd
wave: 4
depends_on: [06-01, 06-02, 06-03]
files_modified:
- src/adapterlab/contracts/evaluation.py
- src/adapterlab/evaluation/plugins.py
- src/adapterl... | adapter-lab | .planning/phases/06-evaluation-reports/06-04-PLAN.md | Markdown | cb9367fcb5e5a025eb75ddbbb62076b6f46125755d0da15bc5882e1bf5885483 | 0 | 896 |
paths, plugin result normalization, review packet generation, annotation import, overlap counts, agreement stats, and privacy scans over plugin and human-review evidence.</action>
<acceptance_criteria>
- `tests/evaluation/test_plugins.py` uses fake entry points for `adapterlab.evaluators`.
- Tests asser... | adapter-lab | .planning/phases/06-evaluation-reports/06-04-PLAN.md | Markdown | 0e030eabd4a386d3cd4f40d4c9aad7f05385cfd07319281d6e94df7911b7c598 | 1 | 896 |
/test_pipeline.py`
- `tests/evaluation/test_manifest.py`
- `src/adapterlab/evaluation/pipeline.py`
- `src/adapterlab/evaluation/manifest.py`
- `src/adapterlab/evaluation/reports.py`
</read_first>
<action>Wire plugin evaluators and human-review evidence into `run_evaluation()`, reports, a... | adapter-lab | .planning/phases/06-evaluation-reports/06-04-PLAN.md | Markdown | 1516dac958e141fd0fd1b959c99b99e1a62ded2b162886bfe30a25f94aa654ab | 2 | 801 |
# 06-04 Summary - Evaluator Plugins, Human Review, and Privacy Tests
**Status:** Complete
**Date:** 2026-05-28
## Delivered
- Added evaluator plugin allowlist evaluation helpers for the `adapterlab.evaluators` entry-point group.
- Added plugin result invocation through the common `EvaluatorResult` contract with plug... | adapter-lab | .planning/phases/06-evaluation-reports/06-04-SUMMARY.md | Markdown | 3411cb0f2b34719312265b7e35404a6538aed71cafa7284131f1aba8a3185c74 | 0 | 213 |
# Phase 6: Evaluation Reports - Context
**Gathered:** 2026-05-28
**Status:** Ready for planning
<domain>
## Phase Boundary
Phase 6 delivers the evaluation subsystem that consumes Phase 5 adapter refs and
`train.manifest.json`, resolves local evaluation suites, runs capability,
multilingual, instruction-following, ag... | adapter-lab | .planning/phases/06-evaluation-reports/06-CONTEXT.md | Markdown | 6656a84df8dd9c183f648c74d697ac33421eae83f9b873772db7c0a4be4ebfc2 | 0 | 896 |
*D-12:** `eval.manifest.json` is the durable evaluation boundary. It records
train manifest refs, adapter/base refs, eval config refs, suite configs,
dataset refs, grader/judge refs, tool inventory refs, seeds, suite versions,
evaluator plugin evidence, environment and library versions, per-suite metric
refs, r... | adapter-lab | .planning/phases/06-evaluation-reports/06-CONTEXT.md | Markdown | 4a9b40bed10f017828d72fa4f7aef9f420f5c95d01349c68d9aa150e1234c554 | 1 | 896 |
dependency on
Phase 5, and four planned Phase 6 plan slices.
- `.planning/STATE.md` - Current milestone position and workflow state.
### Prior Phase Context
- `.planning/phases/01-core-runtime-and-contracts/01-CONTEXT.md` - SDK-first
CLI, strict schemas, manifest envelopes, content-addressed artifacts, plugin
tr... | adapter-lab | .planning/phases/06-evaluation-reports/06-CONTEXT.md | Markdown | e6acc08db5a476f96d95efd69439dcdabd074abeef1e30043b587e94a885f387 | 2 | 896 |
` and `tests/training/test_pipeline.py` -
Existing training manifest and result assertions to extend from the consumer
side.
- `tests/offline/test_no_network.py` - No-network command coverage to extend for
`adapterlab evaluate`.
- `tests/privacy/test_training_outputs.py`,
`tests/privacy/test_synthesis_outputs.p... | adapter-lab | .planning/phases/06-evaluation-reports/06-CONTEXT.md | Markdown | 255364702953f5d30ce5100083b03b55333b6114a26cff85c4b338a8c6e17869 | 3 | 855 |
# Phase 6: Evaluation Reports - Discussion Log
> **Audit trail only.** Do not use as input to planning, research, or execution agents.
> Decisions are captured in CONTEXT.md - this log preserves the alternatives considered.
**Date:** 2026-05-28
**Phase:** 6-Evaluation Reports
**Areas discussed:** Evaluation suite bou... | adapter-lab | .planning/phases/06-evaluation-reports/06-DISCUSSION-LOG.md | Markdown | 5f0c87be6cfd12c488a73e6ab597e3f30c86c870e3b00982032f62cf50aa4278 | 0 | 896 |
to planner discretion.
- Planner may decide whether sanitized per-example evidence is implemented in the first plan or introduced alongside report rendering, as long as reports and manifests remain privacy-safe.
## Deferred Ideas
- Blocking verification gates belong to Phase 7.
- Signing, model cards, bundles, dashbo... | adapter-lab | .planning/phases/06-evaluation-reports/06-DISCUSSION-LOG.md | Markdown | baf9b2bbb44fb18a54fb58612aa81d1a6251c14036b396cb943a027077030636 | 1 | 83 |
# Phase 06 - Pattern Map
## Codebase State
Phase 6 builds on a completed foundation, ingest, synthesis, and adapter training chain:
- `src/adapterlab/cli.py` owns Typer command wiring, `--json` handling, stable error mapping, and `CommandResult` rendering.
- `src/adapterlab/contracts/*` owns strict Pydantic contract... | adapter-lab | .planning/phases/06-evaluation-reports/06-PATTERNS.md | Markdown | c9036ab16a338784c168934eee5c28299e91d99b2e2ab8607cc8a008b16a59af | 0 | 896 |
# Closest Existing Analogs
| New Area | Closest Existing File | Reuse |
|----------|-----------------------|-------|
| Evaluation contracts | `src/adapterlab/contracts/training.py` | Strict config/manifest models, discriminated config variants, artifact refs |
| Config loader | `src/adapterlab/training/config.py` | YA... | adapter-lab | .planning/phases/06-evaluation-reports/06-PATTERNS.md | Markdown | fa7a91d9812c5cc3cf00aee2dfc13aad30e353348b88c9916afbb4e53da4efa0 | 1 | 455 |
# Phase 06: Evaluation Reports - Research
**Researched:** 2026-05-28
**Domain:** Offline evaluation reports, benchmark adapters, evaluator plugins, privacy-safe evidence
**Confidence:** HIGH for adapter-lab architecture and offline guardrails, MEDIUM-HIGH for optional benchmark integration boundaries
<user_constraint... | adapter-lab | .planning/phases/06-evaluation-reports/06-RESEARCH.md | Markdown | 890c25dc91a7edb921a919cec5884621307ef37e6d73d64b2891b6f62adfd018 | 0 | 896 |
remote judges. Sources: XSTest paper/Inspect Evals, HarmBench GitHub, OR-Bench paper. |
| EVL-06 | Tenant custom evaluators can be installed as plugins implementing the evaluator protocol. | Existing plugin registry already knows `adapterlab.evaluators`; Phase 6 should route through `PluginAllowlist` and `PluginRecord`... | adapter-lab | .planning/phases/06-evaluation-reports/06-RESEARCH.md | Markdown | 7db8c22d9597f21ad5965772187318cd326dfe8f74f6806dccb2bf970eece192 | 1 | 896 |
ins and tenant plugins need the same result shape. |
| Optional benchmark adapters | Optional plugins/backends | Evaluation SDK | Heavy libraries must not be base imports. |
| Tool inventory handling | Evaluation SDK | Artifact store | Tool definitions are tenant artifacts and may contain secret-like material. |
| Metr... | adapter-lab | .planning/phases/06-evaluation-reports/06-RESEARCH.md | Markdown | de5a90c6564af62af649207958ab2f9a3dcaf3be2c903dca9b10e915692e87d8 | 2 | 896 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.