ForgeLM good first issues (Python)

TestPatternLinearity still fails on a busy machine: the wall-clock ratio check flakes (42.8x against a 20x limit) 1 day ago

bug good first issue area: dev-tooling severity: low wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

API judge request uses deprecated max_tokens and fixed temperature 0.0, which OpenAI reasoning models reject (CR-A06-051) 2 days ago

bug help wanted area: eval-gates area: integrations dependencies severity: low source: review-2026-10 urgency: next-release wave: 2

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Model-under-test prompts are silently right-truncated at a hard-coded 1024 tokens in both gates: the probe's tail (usually the actual ask) is dropped, and the row still counts as… (CR-A06-011) 2 days ago

bug help wanted area: eval-gates severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

The default DeepSpeed preset zero2.json sets `allgather_bucket_size: "auto"`, which neither transformers nor accelerate fills, so ZeRO-2 aborts (CR-A16-007) 3 days ago

bug good first issue area: packaging-ci severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

notebooks/grpo_reasoning.ipynb: broken example — This config (batch size 1, 4 generations, default gradient accumulation) is a valid GRPO demo config (DOC-D4B05-002) 3 days ago

documentation good first issue severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

forgelm/templates/LICENSES.md: false claim — Templates can rely on the trainer's default for `gradient_checkpointing` (auto-enabled when CUDA is available — see forgelm/trainer.py); only set it expli… (DOC-D4B02-005) 3 days ago

documentation good first issue backlog: known severity: high source: review-2026-10 wave: 1 first-timers-only

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

User-manual quick example fails with exit 1 and prints output the tool never produces (EN and TR) (CR-A12-004) 3 days ago

bug good first issue area: ingestion severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Native review of monitoring-plan and webhook copy in deferred site locales 3 days ago

documentation help wanted area: site i18n severity: low wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

compute_minhash builds a fresh datasketch MinHash (own permutation table) per row, making the 'scale' MinHash path several times slower and heavier than simhash (CR-A11-033) 3 days ago

bug good first issue area: data-audit dependencies severity: medium source: review-2026-10 urgency: next-release wave: 2

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

VRAM probe is untested for MPS and other non-CUDA accelerators (real-gap) (CR-A14-085) 3 days ago

bug good first issue area: onboarding dependencies severity: low source: review-2026-10 urgency: next-release wave: 2

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Quality-filter user manual (EN/TR) opens with a stale heuristic table, a per-row quality_flags list and flag names that do not exist (CR-A11-051) 3 days ago

bug documentation good first issue area: data-audit dependencies severity: medium source: review-2026-10 urgency: next-release wave: 2

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Post-training loaders (--merge, inference.load_model, export merge) ignore model.revision and load the moving default branch (CR-A04-027) 3 days ago

bug good first issue area: pipeline dependencies severity: medium source: review-2026-10 urgency: next-release wave: 2

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Light theme: the hero slider's and the quickstart link's focus rings are 2.26:1 3 days ago

bug good first issue area: site severity: low wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Triage] Declared floor bandit[toml]>=1.7.0 predates the toml extra that the pinned '-c pyproject.toml' invocation relies on (CR-A16-117) 3 days ago

good first issue area: packaging-ci dependencies severity: info source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Triage] Help parsing is sensitive to ANSI colour (latent on Python 3.14): coloured {choices} silently disable choice-value drift checking (CR-A17a-053) 3 days ago

enhancement good first issue area: dev-tooling severity: info source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

.claude/skills/add-test/SKILL.md: broken example — monkeypatch.setattr(forgelm.safety._classifier, "_load_safety_classifier", ...) then run_safety_evaluation(...) intercepts the classifier load (DOC-D4B04-001) 3 days ago

documentation good first issue severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

docs/standards/logging-observability.md: wrong detail — Run JSON output schema: `status` (success|error|reverted|awaiting_approval), `exit_code`, `run_id`, `config_hash`, `metrics`, `resource_usage`,… (DOC-D1B07-003) 3 days ago

documentation good first issue backlog: known severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

cicd.md 'JSON output mode' section documents a non-existent .verdict key and JSON log events on stdout (CR-A04-008) 3 days ago

bug good first issue area: pipeline severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

docs/usermanuals/en/evaluation/benchmarks.md: broken example — The quick example sets `batch_size: 8` under evaluation.benchmark. (DOC-D3B08-019) 3 days ago

documentation good first issue severity: high source: review-2026-10 wave: 1 first-timers-only

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

notebooks/safety_evaluation.ipynb: broken example — results_path = "./safety_checkpoints/final_model/safety_results.json"; if the file is missing, reason 2 is "Model was auto-reverted due to safety f… (DOC-D4B02-006) 3 days ago

documentation good first issue severity: high source: review-2026-10 wave: 1 first-timers-only

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Native review of the German, French, Spanish and Chinese GDPR, retention and masking site strings 3 days ago

help wanted area: site i18n severity: low wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Judge reply parser takes the first JSON block that parses anywhere in the reply, so a score block planted in the model's answer and quoted by the judge replaces the judge's own ve… (CR-A06-015) 5 days ago

bug good first issue area: eval-gates severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Default generative Llama-Guard scorer crashes on every supported transformers 5.x: apply_chat_template returns a BatchEncoding that is passed to generate() as input_ids (CR-A06-003) 5 days ago

bug documentation good first issue area: eval-gates dependencies severity: high source: review-2026-10 urgency: now wave: 0

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Apple Silicon training: measure PyTorch-MPS against MLX and record the backend decision (spike and decision record) 5 days ago

documentation enhancement help wanted area: trainer dependencies severity: info wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

SLERP merge silently returns only the second adapter: state_a aliases base_model storage (CR-A13-001) 6 days ago

bug good first issue area: post-training severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Trend-tracking jq recipe prints evaluation_completed=false as 'n/a', because jq's // operator treats false like null (F-W1-DOCS-TRAIN-16) 7 days ago

documentation good first issue severity: low wave: 4 source: review-2026-09 first-timers-only

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Subparsers re-declare top-level dests (--dry-run, --config, --offline, --output) with plain defaults, so a flag given before the subcommand is silently overwritten (CR-A05-001) 7 days ago

bug good first issue area: cli severity: high source: review-2026-10 wave: 1

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Remove the unreferenced, stale requirements*.txt files (Dependabot will keep bumping them) 7 days ago

enhancement good first issue area: packaging-ci dependencies severity: low wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Numerical-claims guard does not check the stated test-module count outside README 7 days ago

bug good first issue area: dev-tooling severity: low wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Ingestion guide's AWS Textract recipe reads only the first GetDocumentTextDetection page and never checks JobStatus, silently truncating multi-page OCR text (F-W1-DOCS-TRAIN-13) 7 days ago

documentation good first issue severity: medium wave: 3 source: review-2026-09 first-timers-only

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

DPO manual and CLI wizard reverse dpo_beta: they say a lower beta keeps the model closer to the reference model, but in TRL a higher beta does (F-W1-DOCS-TRAIN-07) 7 days ago

documentation good first issue severity: medium wave: 3 source: review-2026-09

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

DeepSpeed platform claims are wrong: doctor says 'ZeRO + offload ... (Linux + macOS only)', and the win32 marker is justified by 'no Windows wheels' when DeepSpeed ships no wheels… (CR-A15-049) 7 days ago

bug good first issue area: operator-cmds dependencies severity: low source: review-2026-10 urgency: next-release wave: 2

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

guide.html pins mermaid@10.9.4, which has published security advisories fixed only in 10.9.6 and 10.9.8 (CR-A18-064) 7 days ago

bug good first issue area: site dependencies severity: low source: review-2026-10 urgency: next-release wave: 2

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

DeepSpeed presets (zero2/zero3/zero3_offload) are not packaged in the wheel or image, so preset names fail on every pip/Docker install (CR-A02-004) 7 days ago

bug documentation help wanted area: trainer backlog: known dependencies severity: high source: review-2026-10 urgency: now wave: 0

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — Docs: docs/usermanuals/en/training/ (part 2/2) 7 days ago

documentation help wanted severity: low source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — Docs: docs/usermanuals/_meta.yaml/ 7 days ago

documentation help wanted severity: low source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — Docs: docs/standards/ (part 2/2) 7 days ago

documentation help wanted severity: low source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — Docs: forgelm (part 2/2) 7 days ago

documentation help wanted severity: low source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — Module sweep: Dev tooling: code, release & supply-chain guards and generators (part 3/3) 7 days ago

bug help wanted area: dev-tooling severity: info source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 3] Medium — Docs: docs/usermanuals/en/data/ (part 2/2) 7 days ago

documentation help wanted severity: medium source: review-2026-10 wave: 3

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 3] Medium — Module sweep: Post-training: inference, chat, GGUF export, deploy, merging (part 2/2) 7 days ago

bug help wanted area: post-training severity: medium source: review-2026-10 wave: 3

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

alignment-overview describes KTO data as (prompt, response, good?), but the KTO loader reads prompt, completion and label 8 days ago

documentation good first issue severity: low wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Browser wizard: the Compliance step throws ReferenceError (appendDetail is not in scope), so risk-assessment, benchmark and judge controls never render (F-W1-SITE-01) 9 days ago

bug good first issue area: site severity: high wave: 1 source: review-2026-09

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 3] Medium — TR mirror: docs/usermanuals/en/training/ 10 days ago

documentation help wanted backlog: known i18n severity: medium source: review-2026-10 wave: 3

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 3] Medium — Docs: docs/usermanuals/tr/deployment/ 10 days ago

documentation help wanted backlog: known i18n severity: medium source: review-2026-10 wave: 3

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 3] Medium — Docs: docs/usermanuals/tr/training/ 10 days ago

documentation help wanted i18n severity: medium source: review-2026-10 wave: 3

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — TR mirror: docs/usermanuals/en/operations/ 10 days ago

documentation help wanted i18n severity: low source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — TR mirror: docs/usermanuals/en/evaluation/ 10 days ago

documentation help wanted i18n severity: low source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — TR mirror: docs/usermanuals/en/data/ (part 2/2) 10 days ago

documentation help wanted i18n severity: low source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — TR mirror: docs/usermanuals/en/concepts/ 10 days ago

documentation help wanted i18n severity: low source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 3] Medium — TR mirror: docs/usermanuals/en/evaluation/ 10 days ago

documentation help wanted i18n severity: medium source: review-2026-10 wave: 3

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Deprecated project.license TOML table + License classifier under an uncapped setuptools build floor (CR-A16-039) 14 days ago

bug help wanted area: packaging-ci backlog: known dependencies severity: low source: review-2026-10 urgency: next-release wave: 2

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Single-source the public-surface mypy gate in `[tool.mypy]` instead of four hand-copied flag strings 14 days ago

help wanted area: dev-tooling severity: low wave: 4 source: roadmap

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Pin batch-size invariance of safety and judge generation in fp32 (`batch_size` 1 / 2 / 4 / 8) 14 days ago

help wanted area: eval-gates severity: low wave: 4 source: roadmap

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Site: the theme-locked --code-comment grey (#6a6f80) on the dark code surface is 3.8:1, below WCAG AA, in both themes (F-W1-SITE-06) 14 days ago

bug good first issue area: site severity: low wave: 4 source: review-2026-09

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Data-audit methods: Presidio documentation URL embedded in error messages (outdated) (TECH-T06-004) 14 days ago

good first issue dependencies source: review-2026-10 urgency: next-release wave: 2

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 3] Medium — Module sweep: Training orchestration, multi-stage pipelines, resume & no-train modes (part 2/2) 14 days ago

bug documentation good first issue area: pipeline backlog: known severity: medium source: review-2026-10 wave: 3

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

[Wave 4] Low/info — Docs: docs/usermanuals/en/concepts/ 14 days ago

documentation good first issue backlog: known severity: low source: review-2026-10 wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Choosing-a-trainer table documents KTO rows as {prompt, response, label}, but the KTO loader requires a completion column (F-W1-DOCS-TRAIN-02) 14 days ago

documentation good first issue severity: medium wave: 3 source: review-2026-09

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

Nightly CI failure — 2026-10-06: flaky TestPatternLinearity timing ratio (21.0x against a 20x limit) 14 days ago

bug good first issue nightly-failure area: dev-tooling severity: low wave: 4

HodeTech/ForgeLM

9

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

Python

#dpo#eu-ai-act#fine-tuning#grpo#huggingface#llm#lora#ml-compliance#peft#trl

HodeTech/ForgeLM

Open Issues Need Help