Environment variables

Every Backpropagate setting can be overridden via an environment variable. Two conventions coexist:

Pydantic-nested-settings shape (BACKPROPAGATE_<GROUP>__<FIELD>, double underscore between group and field) — used for every knob inside the structured Settings model (logging, model, lora, training, data, windows, multirun, security, plus the UI sandbox __OUTPUT_DIR). These are bound via Pydantic’s env_nested_delimiter.
Raw os.environ.get shape (BACKPROPAGATE_<NAME> or BACKPROPAGATE_UI_<NAME>, single underscores throughout) — used for runtime knobs that bypass Pydantic (UI subprocess launch config, debug-traceback toggle, GGUF convert-script discovery, structured-logger config).

Two ways to set them: export in your shell, or put them in a .env file in the working directory (loaded automatically when the [validation] extra is installed).

Logging

Variable	Default	What it does
`BACKPROPAGATE_LOG_LEVEL`	`INFO`	`DEBUG` / `INFO` / `WARNING` / `ERROR`. Sets root logger level.
`BACKPROPAGATE_LOG_JSON`	auto-detect	`true` forces JSON logs, `false` forces console. Auto-detects from TTY by default (JSON when piped, console when interactive).
`BACKPROPAGATE_LOG_FILE`	unset	Path to a file to mirror logs into. If unset, logs go to stderr only.

v1.4 CLI flag overlay: the root parser exposes --log-level, --log-format=json|console, and --log-file=<path> flags that override these env vars for one invocation (CLI flag wins when both are set). See CLI reference → Root-parser flags. The env vars stay the canonical surface for persistent / .env-file-based config.

CLI / runtime opt-outs

Variable	Default	What it does
`BACKPROPAGATE_DEFER_FEATURE_DETECTION`	unset	When set (any truthy value), feature detection at startup is skipped — every optional dep flag stays `False` until you call `refresh_features()` explicitly. Use for the absolute-fastest CLI startup; pays the cost on first real use.
`BACKPROPAGATE_DEFER_FEATURE_QUIET`	unset	Suppresses the one-shot import-time WARN that fires when `BACKPROPAGATE_DEFER_FEATURE_DETECTION` is active. v1.4 Stage C added the WARN so operators aren’t surprised by the deferred-detection auto-refresh path; set this var to `1` if you’ve internalized the trade-off and want quiet startup.
`BACKPROPAGATE_QUIET_TOKEN_HINT`	unset	Suppresses the one-shot stderr note that fires when `cmd_push` falls back to `HF_TOKEN` env var (because no `--token` / `--token-file` was passed). v1.4 Stage C added the note to calibrate the “env var vs file” safety trade-off; set this var to `1` to silence after you’ve calibrated.
`BACKPROPAGATE_ALLOW_CODE_EVAL`	unset	Set to `1` to let the `pass_rate` eval metric run the model’s generated code (same as `backprop eval --allow-code-exec`). Each sample runs in a separate process with a timeout, a stripped environment and resource limits on Linux/macOS, but it is not a sandbox: the code runs as your user. Leave unset unless you would run that code yourself.
`BACKPROPAGATE_DEBUG`	unset	When set to any truthy value, the top-level CLI exception net prints the full Python traceback in addition to the one-line error message. Off by default to keep operator-facing failures short; flip to `1` when filing a bug report.

UI sandbox (FB-003)

Variable	Default	What it does
`BACKPROPAGATE_UI__OUTPUT_DIR`	`~/.backpropagate/ui-outputs`	Single allowed-base directory for all UI-initiated filesystem writes (saved adapters, GGUF exports, converted datasets, Modelfiles). Every UI sink passes this as `allowed_base` to `safe_path` so user-supplied paths cannot escape the sandbox. Denylist-validated — resolves against a denylist of system + credential trees (`/etc`, `/usr`, `/sys`, `/dev`, `/boot`, `/bin`, `/sbin`, `/var/run`, `/var/lib`, `/root`, `~/.ssh`, `~/.aws`, `~/.kube`, `~/.docker`, `~/.gnupg`, `~/.config`, plus the Windows system roots and per-user credential dirs); if the override resolves into a denied path, startup fails with `UI_OUTPUT_DIR_FORBIDDEN`. Bare `/var` is intentionally NOT denied because macOS’s per-user temp tree lives at `/var/folders/<hash>/T/...`; only `/var/run` and `/var/lib` are denied individually. Pick a non-system directory.
`BACKPROPAGATE_UI_PORT`	`7862`	The one port the UI listens on. `backprop ui` runs Reflex in production mode, where the backend serves the compiled frontend, its assets and the `/_event` WebSocket on this port (1.8.0 and earlier also asked for `port + 1`). Read by `cmd_ui` (cli.py), `rxconfig.py` (CORS allowlist) and the auth-badge resolver (ui_security.py).
`BACKPROPAGATE_UI_HOST_BIND`	`127.0.0.1`	Bind host for the UI’s one listener. Read by `cmd_ui` and the auth-badge resolver. Non-loopback overrides (e.g. `0.0.0.0`) require `--auth user:pass` post-v1.2.0; otherwise the runtime refuses to start with `[RUNTIME_UI_AUTH_NOT_ENFORCED]`. SSH port-forwarding remains the lowest-friction remote-access pattern.
`BACKPROPAGATE_UI_SHARE_HOST`	unset	Tunnel host announced to the Host-header allowlist when `--share` is active (cli.py, ui_security.py). v1.3 transitional: populated by the cloudflared tunnel implementation once Wave 6 lands. See the CLI reference.
`BACKPROPAGATE_UI_AUTH`	unset	Plaintext `user:pass` for running Reflex directly (`python -m reflex run` from the package directory); the middleware then enforces HTTP Basic on every HTTP route and the `/_event` WebSocket upgrade. `backprop ui` never uses it: it removes any ambient value and hands the Reflex child `BACKPROPAGATE_UI_AUTH_USER` + `BACKPROPAGATE_UI_AUTH_VERIFIER` instead, so the password stays out of the child’s environment. To keep a password out of shell history, use `--auth-file`.
`BACKPROPAGATE_UI_AUTH_USER`	unset	1.8.1+. HTTP Basic username, set by `backprop ui --auth` / `--auth-file` for the Reflex child. Not a secret. Removed from the ambient environment before every launch.
`BACKPROPAGATE_UI_AUTH_VERIFIER`	unset	1.8.1+. Salted scrypt verifier of the `--auth` password (`scrypt$n$r$p$salt$hash`; n=2^14, r=8, p=1), set by `backprop ui` for the Reflex child. The middleware checks each Basic credential against it in constant time. Takes precedence over `BACKPROPAGATE_UI_AUTH`. A malformed verifier (or one without `BACKPROPAGATE_UI_AUTH_USER`) fails closed: no credential is accepted. Removed from the ambient environment before every launch.
`BACKPROPAGATE_UI_LAUNCH_TOKEN`	unset	Per-launch random token (`secrets.token_urlsafe(32)`). `backprop ui` without `--auth` generates a fresh one on every launch (1.8.1+), prints it in the banner URL (`http://127.0.0.1:<port>/?token=...`) and writes it to a `0600` lock file that is deleted on exit. The first request must carry `?token=`; the middleware answers with a redirect that sets a signed HttpOnly session cookie and drops the token from the address bar. Any ambient value is removed before launch, so a token cannot be pinned.
`BACKPROPAGATE_UI_WORKDIR`	unset	1.8.2+. Where `backprop ui` runs Reflex from. Reflex writes its build and state files (`.web/`, `.states/`, `reflex.lock/`, `uploaded_files/`, a generated `rxconfig.py`) into that folder, so since 1.8.2 it is a per-user folder instead of the installed package: `%LOCALAPPDATA%\backpropagate\ui\<version>-<hash>` on Windows, `$XDG_CACHE_HOME/backpropagate/ui/<version>-<hash>` (default `~/.cache/...`) elsewhere. `<hash>` identifies the install, so a pip install and a Store install keep separate folders; older-version folders of the same install are removed on launch. Set this to use a folder of your own instead, verbatim (no suffix, no clean-up), for example a short path such as `C:\bp-ui` when the default is too deep for Windows’ 260-character path limit.
`BACKPROPAGATE_VRAM_CALIBRATION`	unset	1.8.2+. Path of the file where `backprop estimate-vram --calibrate` stores its measurements. Default: `~/.backpropagate/vram-calibration.json`. One entry per GPU, model, mode and library versions; delete the file to forget every measurement.
`BACKPROPAGATE_UI_PAYLOAD_DIR`	unset	1.8.2+. Location of the prebuilt, offline UI frontend that the Microsoft Store package ships (`ui_frontend_payload/`). When it is present, `backprop ui` checks it (SHA-256), unpacks it into the UI folder once per version and starts with no network access. Normally found next to the installed package; set this only for packaging or testing. A pip install has no payload and is unaffected.
`BACKPROPAGATE_UI_CORS_EXTRA_ORIGINS`	unset	Comma-separated extra CORS origins, additive to the loopback defaults. Used by operators who serve the UI behind a reverse proxy at a non-loopback origin.
`BACKPROPAGATE_UI_QUIET`	unset	When set to `1`, suppresses the 3-line UI startup banner (URL / auth mode / Ctrl+C hint) on stderr. Use for CI / headless launches that don’t want banner noise. Only the exact value `1` suppresses; `0` / `false` / unset leave the banner on.
`BACKPROPAGATE_UI_REQUEST_LOG`	unset	When set to `1` / `true` / `yes` / `on`, enables the v1.3 request-logging ASGI middleware (`ui_app/middleware/request_logging.py`). Emits one `ui.request` structlog line per HTTP request / WebSocket upgrade with fields `method`, `path`, `status`, `duration_ms`, `auth_mode`, `auth_user`, `remote_addr`, `scope_type`. Default OFF — Reflex’s own uvicorn access log covers the basics; opt in when you need structured per-request observability or want to ship records to a SIEM.
`BACKPROPAGATE_UI_RATE_LIMIT_HTTP_PER_MIN`	`100`	v1.3 per-IP cap on rejected HTTP auth attempts (rolling 60s window). Since 1.8.2 only responses the auth middleware rejects (401 / 403 / 421) count, so normal use, page reloads and the UI’s own images never do; before 1.8.2 every request counted and a few reloads made the UI’s icons fail with 429. An IP over the cap gets 429 before its request reaches auth, so brute-force attempts can’t exhaust the HMAC budget. Set to `0` to fully disable HTTP rate-limiting; any positive integer is the per-minute budget. Pass-through paths (`/ping`, `/_next/`, `/favicon`) are never rate-limited.
`BACKPROPAGATE_UI_RATE_LIMIT_WS_PER_MIN`	`10`	v1.3 per-IP cap on rejected WebSocket upgrades (rolling 60s window). Since 1.8.2 only upgrades the auth middleware closes before accepting (4401 / 4403 / 4404) count. An IP over the cap is closed pre-accept with application code `4429`. Set to `0` to fully disable WS rate-limiting.
`BACKPROPAGATE_UI_RATE_LIMIT_UPLOAD_PER_MIN`	`30`	1.8.2 per-IP cap on dataset upload posts (`/_upload`) in a rolling 60 s window; a 31st post in a minute gets 429. `0` disables. Accepted WebSocket frames are never counted.

Security (F-001 / FB-003)

Variable	Default	What it does
`BACKPROPAGATE_SECURITY__REQUIRE_AUTH_FOR_SHARE`	`true`	No-op under the v1.1+ Reflex UI. Held for forward-compat with the Gradio era. Post-v1.2.0 the `--share`-requires-`--auth` contract is enforced by `cli.py:cmd_ui`, not by this flag, and the gate this variable used to relax never fires. The opt-out the variable used to provide is intentionally unavailable — see the CLI reference.
`BACKPROPAGATE_SECURITY__REQUIRE_AUTH`	`false`	Top-level “this is a production deployment” toggle from the Gradio era. Held for forward-compat; the post-v1.2.0 Reflex auth middleware is controlled by `--auth user:pass` and `BACKPROPAGATE_UI_AUTH`, not by this flag.
`BACKPROPAGATE_SECURITY__AUTH_USERNAME`	unset	Auth username for the UI (used by `SecurityConfig.get_auth_tuple()`).
`BACKPROPAGATE_SECURITY__AUTH_PASSWORD`	unset	Auth password for the UI. Stored as a secret in the structured config.
`BACKPROPAGATE_SECURITY__ALLOWED_PATHS`	unset (no restriction)	Comma-separated list of directories the UI is allowed to read/write.
`BACKPROPAGATE_SECURITY__BLOCK_PATH_TRAVERSAL`	`true`	Refuse paths containing `..` segments.
`BACKPROPAGATE_SECURITY__VALIDATE_FILE_MAGIC`	`true`	On by default since 1.8.2. Check an uploaded dataset’s first bytes against its extension (`.jsonl` / `.json` / `.csv` / `.txt` / `.parquet`) before it is stored. `false` trusts the extension alone.
`BACKPROPAGATE_SECURITY__SESSION_TIMEOUT_MINUTES`	`30`	UI session lifetime.
`BACKPROPAGATE_SECURITY__JWT_SECRET`	unset (random)	JWT signing secret. If unset, a random one is generated per process and sessions are lost on restart.
`BACKPROPAGATE_SECURITY__JWT_ALGORITHM`	`HS256`	JWT signing algorithm.
`BACKPROPAGATE_SECURITY__ENABLE_CSRF`	`true`	CSRF protection for state-changing requests.
`BACKPROPAGATE_SECURITY__CSRF_TOKEN_EXPIRY_MINUTES`	`60`	CSRF token lifetime.
`BACKPROPAGATE_SECURITY__RATE_LIMIT_TRAINING`	`3`	Max training-start requests per minute.
`BACKPROPAGATE_SECURITY__RATE_LIMIT_EXPORT`	`5`	Max export requests per minute.
`BACKPROPAGATE_SECURITY__AUDIT_LOG_ENABLED`	`true`	Emit security-audit log lines (rate-limit hits, auth failures, path-traversal blocks).
`BACKPROPAGATE_SECURITY__AUDIT_LOG_FILE`	unset (stdout)	Path for audit log.
`BACKPROPAGATE_SECURITY__ENABLE_CSP`	`true`	Emit a Content-Security-Policy response header.
`BACKPROPAGATE_SECURITY__CSP_REPORT_ONLY`	`false`	When `true`, CSP runs in report-only mode (logs violations, does not block).

Export / GGUF

Variable	Default	What it does
`BACKPROPAGATE_LLAMA_CPP_PATH`	unset	Operator escape hatch for non-standard llama.cpp install locations used by `backprop export --format gguf`. Accepts either the path to `convert_hf_to_gguf.py` directly or the llama.cpp directory containing it. Searched FIRST, before `shutil.which` / `~/llama.cpp` / `/usr/local/bin`. A compiled `llama-quantize` is looked for in the same checkout (the root, `build/bin`, `build/bin/Release`), then on PATH.
`BACKPROPAGATE_UNSLOTH_AUTO_INSTALL`	unset (off)	Set to `1` / `true` / `yes` / `on` to let Unsloth install system packages and build llama.cpp for its own GGUF export. Unsloth does this with `winget` on Windows (apt or brew elsewhere): CMake, compilers and OpenSSL, accepting their licence agreements. Off by default: `import backpropagate` sets `UNSLOTH_AUTO_INSTALL=0`, overriding any value already in the environment, and GGUF export uses the llama.cpp fallback when Unsloth has no built llama.cpp.

cloudflared tunnel (v1.3)

Variable	Default	What it does
`BACKPROPAGATE_CLOUDFLARED_TIMEOUT`	`30` (seconds)	Maximum seconds `backprop ui --share` will wait for `cloudflared` to announce the `https://*.trycloudflare.com` URL on stderr. The default 30s is comfortable on a healthy uplink; operators on slow uplinks can extend the budget. On timeout the CLI emits a clear error pointing at SSH port-forwarding as the fallback.

Model

Variable	Default	What it does
`BACKPROPAGATE_MODEL__NAME`	`Qwen/Qwen2.5-7B-Instruct`	HuggingFace model id or local path. This is the canonical default when `Trainer()` is called with no model argument.
`BACKPROPAGATE_MODEL__LOAD_IN_4BIT`	`true`	4-bit quantization at load time (saves ~50% VRAM).
`BACKPROPAGATE_MODEL__MAX_SEQ_LENGTH`	`2048`	Maximum sequence length.
`BACKPROPAGATE_MODEL__DTYPE`	unset (auto)	Force `bf16` / `fp16` / `fp32`. Auto-detects bf16 on Ampere+.
`BACKPROPAGATE_MODEL__TRUST_REMOTE_CODE`	`false`	Whether to run custom Python code that a model’s Hugging Face repository ships. Off by default since v1.7.2: a model that needs it fails with `CONFIG_TRUST_REMOTE_CODE_REQUIRED`. Set to `true` only for a repository you have read and trust. None of the curated presets need it.

LoRA

Variable	Default	What it does
`BACKPROPAGATE_LORA__R`	`256`	LoRA rank. Changing it (or the alpha or target modules below) turns the automatic LoRA preset choice off: your values are used as given.
`BACKPROPAGATE_LORA__LORA_ALPHA`	`512`	LoRA scaling factor (alpha = 2 x rank by convention).
`BACKPROPAGATE_LORA__LORA_DROPOUT`	`0.05`	LoRA dropout rate.
`BACKPROPAGATE_LORA__USE_GRADIENT_CHECKPOINTING`	`unsloth`	`unsloth` / `true` / `false`.
`BACKPROPAGATE_LORA__RANDOM_STATE`	`42`	RNG seed for LoRA init reproducibility.
`BACKPROPAGATE_LORA__USE_RSLORA`	`false`	v1.5 — rank-stabilized LoRA scaling (`alpha/sqrt(r)` instead of `alpha/r`). Zero inference cost, still mergeable; the benefit grows with rank (relevant at the rank-256 default). Equivalent to `--use-rslora` on the CLI.

Training

Variable	Default	What it does
`BACKPROPAGATE_TRAINING__PER_DEVICE_TRAIN_BATCH_SIZE`	`2`	Per-device batch size.
`BACKPROPAGATE_TRAINING__GRADIENT_ACCUMULATION_STEPS`	`4`	Effective batch size = batch × grad-accum.
`BACKPROPAGATE_TRAINING__MAX_STEPS`	`100`	Hard cap on training steps.
`BACKPROPAGATE_TRAINING__NUM_TRAIN_EPOCHS`	`1`	Number of epochs (ignored if `MAX_STEPS > 0`).
`BACKPROPAGATE_TRAINING__LEARNING_RATE`	`2e-4`	Learning rate.
`BACKPROPAGATE_TRAINING__WEIGHT_DECAY`	`0.01`	Weight decay.
`BACKPROPAGATE_TRAINING__WARMUP_STEPS`	`10`	Number of warmup steps.
`BACKPROPAGATE_TRAINING__WARMUP_RATIO`	`0.0`	Warmup ratio (alternative to warmup steps).
`BACKPROPAGATE_TRAINING__OPTIM`	`adamw_8bit`	Optimizer name.
`BACKPROPAGATE_TRAINING__LR_SCHEDULER_TYPE`	`cosine`	LR schedule.
`BACKPROPAGATE_TRAINING__LOGGING_STEPS`	`10`	Log every N steps.
`BACKPROPAGATE_TRAINING__SAVE_STEPS`	`100`	Save a checkpoint every N steps.
`BACKPROPAGATE_TRAINING__BF16`	`true`	Use bf16 (Ampere+ recommended).
`BACKPROPAGATE_TRAINING__FP16`	`false`	Use fp16 (older GPUs).
`BACKPROPAGATE_TRAINING__SEED`	`42`	RNG seed.
`BACKPROPAGATE_TRAINING__OUTPUT_DIR`	`./output`	Where to write checkpoints and exports.
`BACKPROPAGATE_TRAINING__OVERWRITE_OUTPUT_DIR`	`true`	Whether to overwrite an existing output dir.
`BACKPROPAGATE_TRAINING__METHOD`	`sft`	v1.5 (orpo) / v1.6 (simpo, kto) — training objective. One of `sft` / `orpo` / `simpo` / `kto`. `sft` (default) = supervised fine-tuning. `orpo` and `simpo` = reference-free preference tuning (need a paired `{prompt, chosen, rejected}` dataset; single-stage, no reference model). `kto` = Kahneman-Tversky Optimization on unpaired `{prompt, completion, label}` binary-feedback data (LoRA-only). Any other value is rejected with `CONFIG_INVALID_SETTING`. See Preference tuning.
`BACKPROPAGATE_TRAINING__ORPO_BETA`	`0.1`	v1.5 — ORPO odds-ratio weight (lambda). Keep > 0. Ignored unless `METHOD=orpo`.
`BACKPROPAGATE_TRAINING__SIMPO_BETA`	`2.0`	v1.6 — SimPO reward-scaling temperature (`beta` in TRL’s `CPOConfig` under `loss_type="simpo"`). The cross-setup safe floor from the SimPO paper (arXiv:2405.14734). Any finite value is admissible; the `gamma/beta` ratio is what carries the degeneration risk (warned, not blocked). Ignored unless `METHOD=simpo`.
`BACKPROPAGATE_TRAINING__SIMPO_GAMMA`	`1.0`	v1.6 — SimPO target reward margin (`gamma`; absolute, = `beta`×0.5 at the default beta). Must be > 0 — a non-positive margin degenerates the objective (`CONFIG_INVALID_SETTING`). A `gamma/beta` ratio above 1.0 risks repetitive output and emits a WARN (not an error). Ignored unless `METHOD=simpo`.
`BACKPROPAGATE_TRAINING__KTO_BETA`	`0.1`	v1.6 — KTO loss temperature (`beta` in TRL’s `KTOConfig`; the KTO paper / TRL default, arXiv:2402.01306). Ignored unless `METHOD=kto`.
`BACKPROPAGATE_TRAINING__KTO_DESIRABLE_WEIGHT`	`1.0`	v1.6 — loss weight on desirable (`label=true`) KTO examples. Must be > 0 (`CONFIG_INVALID_SETTING`). This is the operator’s starting point: the trainer auto-rebalances the effective desirable/undesirable weights from your label counts toward the `[1:1, 4:3]` band, so the final ratio is data-derived, not this raw value. Ignored unless `METHOD=kto`.
`BACKPROPAGATE_TRAINING__KTO_UNDESIRABLE_WEIGHT`	`1.0`	v1.6 — loss weight on undesirable (`label=false`) KTO examples. Must be > 0 (`CONFIG_INVALID_SETTING`). Auto-rebalanced from label counts alongside `KTO_DESIRABLE_WEIGHT` (see above). Ignored unless `METHOD=kto`.
`BACKPROPAGATE_TRAINING__FP8`	`false`	v1.5; verified on Blackwell in v1.6 — FP8 compute path on Blackwell (RTX 5090, sm_120) / Hopper (sm_90+) via torchao. Base projection weights in float8 (~1.4x throughput, ~60% less base memory); the LoRA adapter stays bf16 and the merge → GGUF → Ollama export still works. `mode='lora'` + `method='sft'` only; falls back to bf16 with a warning if unsupported (a broken torchao install raises `RUNTIME_FP8_UNSUPPORTED`). Needs `pip install 'backpropagate[fp8]'`. Equivalent to `--fp8` on the CLI.
`BACKPROPAGATE_TRAINING__BACKEND`	`auto`	v1.5 T3.1 (experimental — Apple-Silicon rail BUILT-BUT-UNVERIFIED) — the compute rail. One of `auto` / `cuda` / `mlx`. `auto` (default) routes to CUDA on an NVIDIA host and to the MLX (`mlx_lm.lora`) rail on an Apple-Silicon Mac with the `[mlx]` extra — existing CUDA rigs stay byte-identical. `cuda` forces the CUDA rail; `mlx` forces the Apple-Silicon rail. MLX is LoRA SFT only in v1.5 (`method='orpo'` / `mode='full'` / `fp8=True` / multi-run rejected with `CONFIG_INVALID_SETTING`). A forced `mlx` on a non-Apple host errors with `CONFIG_INVALID_SETTING`; if the resolved rail is `mlx` but `mlx_lm` is missing, the run raises `DEP_MLX_UNAVAILABLE` (`pip install 'backpropagate[mlx]'`). Equivalent to `--backend` on the CLI. The MLX rail is built + unit-tested (mocked), pending dogfood verification on real Apple Silicon.

Full fine-tuning offload (`--full-ft-offload`)

Variable	Default	What it does
`BACKPROPAGATE_OFFLOAD_PIN`	`register`	How the offload engine page-locks host memory. `register` page-locks the parameter storage in place: no copy, exact size. `pinned` uses PyTorch’s pinned allocator, which made steps 3-5x faster in our runs but rounds every block up to a power of two; at 7B that pushed host RAM past 60 GiB. `none` uses ordinary pageable memory: slowest, smallest.
`BACKPROPAGATE_OFFLOAD_ROUNDING`	`stochastic`	Diagnostic only. The engine keeps weights in bf16 and writes each update back with stochastic rounding. `nearest` switches to round-to-nearest, which drops most small updates, so the run stops learning. It exists to test the engine’s own safety check. Leave it at the default.

Data

Variable	Default	What it does
`BACKPROPAGATE_DATA__DATASET_NAME`	`HuggingFaceH4/ultrachat_200k`	Default HF dataset when none is passed.
`BACKPROPAGATE_DATA__DATASET_SPLIT`	`train_sft`	Which split to load.
`BACKPROPAGATE_DATA__MAX_SAMPLES`	`0`	Cap the training set to N randomly chosen rows. `0` (the default) uses every row. Before v1.7.2 the default was `1000`, which silently dropped the rest of any larger dataset.
`BACKPROPAGATE_DATA__TEXT_COLUMN`	`text`	Column name for raw-text datasets.
`BACKPROPAGATE_DATA__CHAT_FORMAT`	`chatml`	Chat template (`chatml` / `llama` / `alpaca` / `sharegpt`).
`BACKPROPAGATE_DATA__PRE_TOKENIZE`	`true`	Pre-tokenize before training (Windows-safe).
`BACKPROPAGATE_DATA__SHUFFLE`	`true`	Shuffle the dataset.
`BACKPROPAGATE_DATA__PACKING`	`true`	Combine short sequences via TRL sample packing. v1.3 default flipped from `false` to `true` per BACKEND-4 — 1.7-3× wall-clock throughput on SFT runs. Set to `false` to opt out (boundary-token leakage with exotic chat templates is the documented edge case).
`BACKPROPAGATE_DATA__REASONING_TRACE`	`false`	v1.5 T3.2 — reasoning-trace SFT (R1/QwQ distillation). When `true` the trainer keeps the `<think>` chain-of-thought in the SFT target, applies trace-length filtering (drops empty / over-long traces), and raises the default `max_seq_length` to `8192` if it was left at the shipped `2048`. `<think>` is plain text — no special tokens, no embedding resize — so the merge→GGUF→Ollama export is unaffected. SFT only — ignored under any preference method (`orpo` / `simpo` / `kto`), with a one-line advisory if you set both. Default `false` = byte-identical v1.4 SFT. Equivalent to `--reasoning-trace` on the CLI.
`BACKPROPAGATE_DATA__MIN_TRACE_TOKENS`	`8`	Minimum `<think>` token count to keep a sample under `reasoning_trace`. Samples whose reasoning span tokenizes below this are dropped as empty / degenerate traces.
`BACKPROPAGATE_DATA__MAX_TRACE_TOKENS`	`8192`	Maximum `<think>` token count to keep a sample under `reasoning_trace`. Samples whose reasoning span tokenizes above this are dropped as over-long traces.

Windows

Variable	Default	What it does
`BACKPROPAGATE_WINDOWS__DATALOADER_NUM_WORKERS`	`0`	Workers for the DataLoader (0 is the safe default on Windows).
`BACKPROPAGATE_WINDOWS__TOKENIZERS_PARALLELISM`	`false`	Forwarded to HF tokenizers.
`BACKPROPAGATE_WINDOWS__XFORMERS_DISABLED`	`true`	Disable xformers (incompatible with SM 12.0+).
`BACKPROPAGATE_WINDOWS__CUDA_LAUNCH_BLOCKING`	`false`	Set `CUDA_LAUNCH_BLOCKING=1`; useful for debugging, slows training.
`BACKPROPAGATE_WINDOWS__PRE_TOKENIZE`	`true`	Pre-tokenize to avoid multiprocessing issues.

Multi-run

Variable	Default	What it does
`BACKPROPAGATE_MULTIRUN__NUM_RUNS`	`5`	Number of training runs in a multi-run campaign.
`BACKPROPAGATE_MULTIRUN__STEPS_PER_RUN`	`100`	Steps per run.
`BACKPROPAGATE_MULTIRUN__SAMPLES_PER_RUN`	`1000`	Samples per run.
`BACKPROPAGATE_MULTIRUN__CONTINUE_FROM_PREVIOUS`	`true`	Resume from the previous LoRA each run.
`BACKPROPAGATE_MULTIRUN__SAVE_INTERMEDIATE`	`true`	Save intermediate checkpoints.