MuhibBeekun

pickai-qwen2.5-7b-nl-parse-lora

Overview

PickAI separates constraint parsing from route optimization. A language model (optional) turns operator instructions into JSON constraints. OR-Tools or a heuristic engine computes the actual pick sequence. This repository holds the PEFT LoRA weights from a bounded fine-tune experiment on synthetic warehouse NL data.

mermaid

flowchart LR  NL[Supervisor NL] --> Parser[NL parser]  Parser --> JSON[OptimizeConstraints JSON]  JSON --> Solver[OR-Tools / heuristic]  Wave[Released wave lines] --> Solver  Solver --> Route[Optimized pick sequence]

What this adapter targets

Table with columns: Field, Type, Example intent
Field	Type	Example intent
`equipment_mode`	`walker` \| `forklift`	"Use forklift mode for this wave"
`ladder_must_stay_in_aisle`	boolean	"Keep the ladder in aisle 3"
`start_position`	`{ aisle, level, x, y }`	"Start at aisle B, level 2"
`depot`	`{ aisle, level, x, y }`	Optional depot override

Full schema: PickAI contracts.

Recommended use

Table with columns: Use case, Recommendation
Use case	Recommendation
PickAI production / Docker default	Base Qwen via Ollama (99.33% held-out aggregate)
Research on warehouse NL → structured constraints	Load this adapter with PEFT for experimentation
Route optimization	Do not use this adapter — use PickAI's deterministic solver

Training

Table with columns: Setting, Value
Setting	Value
Base model	Qwen/Qwen2.5-7B-Instruct
Method	PEFT LoRA
Rank (`r`)	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	`q_proj`, `k_proj`, `v_proj`,

Training loss (logged every 10 steps)

Table with columns: Step, Loss
Step	Loss
10	2.155
20	0.220
30	0.103
40	0.038
50	0.026
60	0.022
70	0.018
80	0.018
90	0.017
100

Training loss dropped sharply while held-out field accuracy collapsed — a classic overfit / train-inference mismatch signal.

Training script: scripts/train_lora_nl_parse.py

Evaluation

100 held-out examples (deterministic hash split). Scoring: exact field match on equipment_mode, start_position, and ladder_must_stay_in_aisle.

Parity pass (prompt-aligned train + eval, June 2026)

Table with columns: Metric, Base (Ollama), This LoRA
Metric	Base (Ollama)	This LoRA
Aggregate field match	100.00%	44.67%
Equipment mode	100.00%	28.00%
Ladder position	100.00%	28.00%
Aisle constraint	100.00%	78.00%

First run (misaligned train prompt, tail holdout)

Table with columns: Metric, Base (Ollama), This LoRA
Metric	Base (Ollama)	This LoRA
Aggregate field match	99.33%	17.67%

V2 pass (500 steps, rank 32, June 2026)

Table with columns: Metric, Base (Ollama), LoRA v2 (local)
Metric	Base (Ollama)	LoRA v2 (local)
Aggregate field match	99.33%	16.67%
Equipment mode	98.00%	0.00%
Ladder position	100.00%	0.00%
Aisle constraint	100.00%	50.00%

More training capacity did not help; release runtime stays on base Qwen.

Value gate: failed. PickAI does not enable this adapter by default. Parity fixes improved LoRA from 17.67% to 44.67% aggregate but base Qwen still wins.

Full write-up: docs/fine-tune-eval.md

Likely causes (working hypotheses)

Base Qwen via Ollama was already near ceiling before training
First run used a different train prompt than eval; parity pass corrected that gap partially
Synthetic phrasing still overfits; ladder start_position remains the weakest field

Load locally (PEFT)

Requires transformers, peft, torch, and the base model weights from Hugging Face.

python

import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = "Qwen/Qwen2.5-7B-Instruct"adapter = "MuhibBeekun/pickai-qwen2.5-7b-nl-parse-lora" tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(    base,    torch_dtype=torch.float16,    device_map="auto",    trust_remote_code=True,)model = PeftModel.from_pretrained(model, adapter)model.eval()

Example prompt format (matches training and eval):

python

from pickai.inference.nl_parse_prompt import build_nl_parse_prompt prompt = build_nl_parse_prompt(    "Use forklift mode and keep the ladder in aisle 3.",    {"orders": 7, "lines": 12, "x_min": 2.0, "x_max": 24.0, "y_min": 7.0, "y_max": 42.0},)

PickAI integration (optional)

PickAI ships with base Ollama parsing. To experiment with this adapter inside the repo:

powershell

git clone https://github.com/Muhib-Beekun/pickai.gitcd pickai# train or download adapter into outputs/lora$env:PICKAI_USE_LORA = "1"$env:PICKAI_LOCAL_LORA_DIR = "outputs/lora"docker compose up -d --build

Re-evaluate before relying on it:

powershell

python scripts/eval_nl_parse.py --lora-path outputs/lora

Table with columns: Resource, URL
Resource	URL
PickAI (GitHub)	https://github.com/Muhib-Beekun/pickai
Training dataset	https://huggingface.co/datasets/MuhibBeekun/pickai-synthetic-nl-parse-v1
Fine-tune evaluation	https://github.com/Muhib-Beekun/pickai/blob/main/docs/fine-tune-eval.md
WMS integration guide	https://github.com/Muhib-Beekun/pickai/blob/main/docs/wms-integration-guide.md
Upstream inspiration

License

MIT. See PickAI LICENSE and upstream picking-route attribution in NOTICE.md.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

MuhibBeekun

Model Tree

Base

Qwen/Qwen2.5-7B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

Overview

mermaid

flowchart LR  NL[Supervisor NL] --> Parser[NL parser]  Parser --> JSON[OptimizeConstraints JSON]  JSON --> Solver[OR-Tools / heuristic]  Wave[Released wave lines] --> Solver  Solver --> Route[Optimized pick sequence]

flowchart LR  NL[Supervisor NL] --> Parser[NL parser]  Parser --> JSON[OptimizeConstraints JSON]  JSON --> Solver[OR-Tools / heuristic]  Wave[Released wave lines] --> Solver  Solver --> Route[Optimized pick sequence]

What this adapter targets

Table with columns: Field, Type, Example intent
Field	Type	Example intent
`equipment_mode`	`walker` \| `forklift`	"Use forklift mode for this wave"
`ladder_must_stay_in_aisle`	boolean	"Keep the ladder in aisle 3"
`start_position`	`{ aisle, level, x, y }`	"Start at aisle B, level 2"
`depot`	`{ aisle, level, x, y }`	Optional depot override

Full schema: PickAI contracts.

Recommended use

Table with columns: Use case, Recommendation
Use case	Recommendation
PickAI production / Docker default	Base Qwen via Ollama (99.33% held-out aggregate)
Research on warehouse NL → structured constraints	Load this adapter with PEFT for experimentation
Route optimization	Do not use this adapter — use PickAI's deterministic solver

Training

Table with columns: Setting, Value
Setting	Value
Base model	Qwen/Qwen2.5-7B-Instruct
Method	PEFT LoRA
Rank (`r`)	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	`q_proj`, `k_proj`, `v_proj`,

Training loss (logged every 10 steps)

Table with columns: Step, Loss
Step	Loss
10	2.155
20	0.220
30	0.103
40	0.038
50	0.026
60	0.022
70	0.018
80	0.018
90	0.017
100

Training loss dropped sharply while held-out field accuracy collapsed — a classic overfit / train-inference mismatch signal.

Training script: scripts/train_lora_nl_parse.py

Evaluation

100 held-out examples (deterministic hash split). Scoring: exact field match on equipment_mode, start_position, and ladder_must_stay_in_aisle.

Parity pass (prompt-aligned train + eval, June 2026)

Table with columns: Metric, Base (Ollama), This LoRA
Metric	Base (Ollama)	This LoRA
Aggregate field match	100.00%	44.67%
Equipment mode	100.00%	28.00%
Ladder position	100.00%	28.00%
Aisle constraint	100.00%	78.00%

First run (misaligned train prompt, tail holdout)

Table with columns: Metric, Base (Ollama), This LoRA
Metric	Base (Ollama)	This LoRA
Aggregate field match	99.33%	17.67%

V2 pass (500 steps, rank 32, June 2026)

Table with columns: Metric, Base (Ollama), LoRA v2 (local)
Metric	Base (Ollama)	LoRA v2 (local)
Aggregate field match	99.33%	16.67%
Equipment mode	98.00%	0.00%
Ladder position	100.00%	0.00%
Aisle constraint	100.00%	50.00%

More training capacity did not help; release runtime stays on base Qwen.

Value gate: failed. PickAI does not enable this adapter by default. Parity fixes improved LoRA from 17.67% to 44.67% aggregate but base Qwen still wins.

Full write-up: docs/fine-tune-eval.md

Likely causes (working hypotheses)

Base Qwen via Ollama was already near ceiling before training
First run used a different train prompt than eval; parity pass corrected that gap partially
Synthetic phrasing still overfits; ladder start_position remains the weakest field

Load locally (PEFT)

Requires transformers, peft, torch, and the base model weights from Hugging Face.

python

import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = "Qwen/Qwen2.5-7B-Instruct"adapter = "MuhibBeekun/pickai-qwen2.5-7b-nl-parse-lora" tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(    base,    torch_dtype=torch.float16,    device_map="auto",    trust_remote_code=True,)model = PeftModel.from_pretrained(model, adapter)model.eval()

Example prompt format (matches training and eval):

python

from pickai.inference.nl_parse_prompt import build_nl_parse_prompt prompt = build_nl_parse_prompt(    "Use forklift mode and keep the ladder in aisle 3.",    {"orders": 7, "lines": 12, "x_min": 2.0, "x_max": 24.0, "y_min": 7.0, "y_max": 42.0},)

PickAI integration (optional)

PickAI ships with base Ollama parsing. To experiment with this adapter inside the repo:

powershell

git clone https://github.com/Muhib-Beekun/pickai.gitcd pickai# train or download adapter into outputs/lora$env:PICKAI_USE_LORA = "1"$env:PICKAI_LOCAL_LORA_DIR = "outputs/lora"docker compose up -d --build

git clone https://github.com/Muhib-Beekun/pickai.gitcd pickai# train or download adapter into outputs/lora$env:PICKAI_USE_LORA = "1"$env:PICKAI_LOCAL_LORA_DIR = "outputs/lora"docker compose up -d --build

Re-evaluate before relying on it:

powershell

python scripts/eval_nl_parse.py --lora-path outputs/lora

python scripts/eval_nl_parse.py --lora-path outputs/lora

Table with columns: Resource, URL
Resource	URL
PickAI (GitHub)	https://github.com/Muhib-Beekun/pickai
Training dataset	https://huggingface.co/datasets/MuhibBeekun/pickai-synthetic-nl-parse-v1
Fine-tune evaluation	https://github.com/Muhib-Beekun/pickai/blob/main/docs/fine-tune-eval.md
WMS integration guide	https://github.com/Muhib-Beekun/pickai/blob/main/docs/wms-integration-guide.md
Upstream inspiration

License

MIT. See PickAI LICENSE and upstream picking-route attribution in NOTICE.md.

pickai-qwen2.5-7b-nl-parse-lora

README

Overview

Recommended use

Training

Evaluation

Load locally (PEFT)

PickAI integration (optional)

License

Explore FriendliAI today

README

Overview

Recommended use

Training

Evaluation

Load locally (PEFT)

PickAI integration (optional)

License

pickai-qwen2.5-7b-nl-parse-lora

README

Overview

Recommended use

Training

Evaluation

Load locally (PEFT)

PickAI integration (optional)

Related links

License

Explore FriendliAI today

README

Overview

Recommended use

Training

Evaluation

Load locally (PEFT)

PickAI integration (optional)

Related links

License