FR-Forge-1.7B

A Thin Language Model for manufacturing. Grounded, specialized, and running on your laptop.

GROUNDED.   SPECIALIZED.   LOCAL.

---

Contents

1. Overview
2. How it works
3. Specifications
4. Quickstart
5. Intended use
6. Evaluation
7. Limitations
8. Training
9. Citation

Overview

FR-Forge-1.7B is a Thin Language Model (TLM) for the manufacturing sector, by Fahrenheit Research. Sibling to FR-Lex (legal). It is a small, 4-bit, MLX model fine-tuned with LoRA adapters on a curated manufacturing instruction corpus, designed to run locally on Apple Silicon.

  Where a Thin Language Model fits

  specialization
     ▲
     │   ● FR-Forge 1.7B
     │     narrow · local · cheap
     │
     │                              ● Frontier LLM
     │                                broad · hosted · costly
     └──────────────────────────────────────────────▶ generality

How it works

  Manufacturing question
            │
            ▼
  ┌─────────────────────────────────────────────────────┐
  │ FR-Forge 1.7B                                         │
  └─────────────────────────────────────────────────────┘
            │
   reasons across three pillars
            │
            ├─ Ops & Maintenance        SOPs · PM/PdM · OEE · troubleshooting
            ├─ Quality & Compliance     ISO 9001 · IATF 16949 · FDA/GMP · CAPA · FMEA · SPC
            └─ Supply Chain & Planning  MRP/ERP · BOMs · inventory · demand planning
            │
            ▼
   Grounded, local answer  (verify exact clauses against the controlling standard)

It is an assistant, not a certified authority. It is not a substitute for the controlling standard, safety sign-off, or regulatory advice.

Specifications

Base	Qwen3-1.7B (4-bit, MLX)
Parameters	1.7B · 4-bit
Method	LoRA adapters, fused
Runtime	MLX (Apple Silicon)
Languages	English
License	Apache-2.0

Quickstart

Runs locally with MLX on Apple Silicon. Apply a repetition penalty to avoid looping.

pip install -U mlx-lm
python3 -m mlx_lm generate \
  --model FahrenheitResearch/FR-Forge-1.7B \
  --repetition-penalty 1.15 \
  --max-tokens 500 \
  --prompt "How should I set safety stock for a long-lead component with variable demand?"

Recommended generation settings: repetition_penalty = 1.15 (raise to 1.3 if output repeats). A system prompt of "You are FR-Forge, a manufacturing domain assistant." matches training.

Intended use

Assisting manufacturing teams with everyday domain questions across three pillars:

  FR-Forge covers
  ├─ Shop-floor ops & maintenance   SOPs, work instructions, equipment manuals, PM/PdM, OEE, troubleshooting
  ├─ Quality & compliance           ISO 9001, IATF 16949, ISO 13485, FDA/GMP, CAPA, FMEA, SPC, MSA, audits
  └─ Supply chain & planning         MRP/ERP, procurement, BOMs, inventory, demand planning, suppliers

Out of scope. Safety, regulatory, or compliance sign-off; anything that must be exact (part numbers, clause text, customer-specific requirements) without grounding it in your own documents.

Evaluation

A held-out set of prompts across the three pillars is scored with a deterministic keyword-rubric: each item defines groups of required terms, and a group passes if any synonym appears in the answer. The item score is the fraction of groups covered; pillar and overall scores are averages. Generation uses a 1.15 repetition penalty.

  Held-out eval, keyword rubric (percent coverage)

  Ops & Maintenance        ███████████████████░  92.9
  Quality & Compliance     ██████████████░░░░░░  70.8
  Supply Chain & Planning  █████████████████░░░  83.3
  ───────────────────────  ────────────────────  ────
  Overall                  ████████████████░░░░  82.3

Pillar	Score
Ops & Maintenance	92.9%
Quality & Compliance	70.8%
Supply Chain & Planning	83.3%
Overall	82.3%

This measures domain-term coverage, not eloquence or factual grading, and the held-out set is small, so treat results as a directional, reproducible yardstick rather than a precise grade. The untuned base can be scored with python3 scripts/evaluate.py --base-only, and the tuned model is reproduced with python3 scripts/evaluate.py --max-tokens 500.

Limitations

Not a certified authority. Outputs assist research and drafting only. They are not a substitute for the controlling standard, safety sign-off, or regulatory advice.

• English only (v1).
• Paraphrased, not verbatim. Trained on domain reasoning, not reproduced standards; always verify clause-level detail against the controlling standard.
• Small model. For facts that must be exact (part numbers, clause text, customer-specific requirements), ground it with retrieval over your own documents rather than relying on memory.
• Self-reported metric. Evaluation is an internal keyword-coverage score on a small held-out set.

Training

  Domain instruction pairs ─┐
  110 ops & maintenance      │
  100 quality & compliance   ├─▶ LoRA fine-tune (MLX) ─▶ fuse adapters ─▶ FR-Forge 1.7B
   51 supply chain & planning┘    iters 800 · lr 1e-5 · seq 512

  Training data, 261 instruction pairs
  Ops & Maintenance        ████████████████  110
  Quality & Compliance     ███████████████░  100
  Supply Chain & Planning  ███████░░░░░░░░░   51

Base, method, data, and hyperparameters

• Base model: mlx-community/Qwen3-1.7B-4bit (Qwen3 architecture, 4-bit)
• Method: LoRA adapters via mlx-lm, fused into this standalone model
• Data: 261 curated instruction pairs (110 ops & maintenance, 100 quality & compliance, 51 supply chain & planning). Sources: paraphrased domain reasoning plus model-assisted synthetic pairs with human review. No copyrighted standard text is reproduced verbatim; the data teaches structure and reasoning and cites clause numbers.
• Hyperparameters: iters 800, LoRA layers 8, batch size 1, max sequence length 512, learning rate 1e-5, gradient checkpointing on. Peak training memory ~2 GB.

Citation

Apache-2.0. Base model mlx-community/Qwen3-1.7B-4bit is Apache-2.0.

@software{fr_forge_2026,
  title  = {FR-Forge-1.7B: a thin language model for manufacturing},
  author = {Fahrenheit Research},
  year   = {2026},
  note   = {Fine-tuned from Qwen3-1.7B (4-bit) with MLX/LoRA}
}

---

FAHRENHEIT RESEARCH

Thin Language Models for specialized domains.

Website  ·  GitHub  ·  Sibling model: FR-Lex-1.7B