SkyIsNotGreen/hive-dispatcher-1.7b-v1

Hive Dispatcher 1.7B

The Hive 1.7B roster

This model is one seat in the five-model Hive 1.7B family. They are role specialists, not five general chat models:

model	seats
Hive Conductor 1.7B	controller + planner
Hive Reader 1.7B	mechanics + RAG
Hive Auditor 1.7B	placement + critic
Hive Dispatcher 1.7B	dispatcher
Hive Tool Manager 1.7B	agency

The family is served and routed by Hive.

A dispatcher specialist for the Hive local control plane: decide which worker runs each job, in what order, and whether the whole plan fits the deadline.

Built from LiquidAI/LFM3-1.7B-Base, this specialist covers the dispatcher seat. It is not a general chat model.

What it is

The dispatcher seat is the "who does what, when" layer. It handles four task families, all evaluated on held-out tasks:

family	tasks	what it does
job assignment	12	pack jobs with durations onto workers with time budgets; assign every job exactly once without exceeding a worker's capacity
dependency scheduling	15	place jobs on worker timelines with precedence constraints; return worker and start time for each job so the makespan meets a deadline
validation	9	check a proposed dispatch against worker capacities and answer yes/no
priority triage	4	when two jobs need the only free worker, choose the one with the nearer deadline (for example, a live rollback before a background re-index)

Example

A dispatcher task gives the jobs, workers, durations, dependencies and a deadline. The model returns the plan, not prose.

Input:

Schedule jobs A(5), B(1), C(4), D(5), E(2) on workers W1, W2, W3. Dependencies: A before B; B before C; B before D; A before E. Makespan must be at most 12.

Output:

{"schedule": [
  {"task": "A", "worker": "W1", "start": 0},
  {"task": "B", "worker": "W1", "start": 5},
  {"task": "C", "worker": "W2", "start": 6},
  {"task": "D", "worker": "W1", "start": 6},
  {"task": "E", "worker": "W3", "start": 5}
]}

The plan puts A, B and D on Worker 1, C on Worker 2 and E on Worker 3. It finishes at minute 11, inside the 12-minute deadline. The model does not run the jobs or control the workers; it returns the plan for the caller to apply.

The other output shapes are:

task	output
assignment	{"assignments":[{"task":"A","worker":"W1"}]}
validation	yes or no
priority triage	A or B

The suite's task format uses W1, W2, W3 worker IDs; the hero spells them out as Worker 1/2/3.

The hero

The hero shows one dependency-scheduling pass (times in minutes):

• Worker 1: BUILD 0-4, TEST 4-7, RELEASE 7-12
• Worker 2: WRITE 0-4 (write the release notes)
• Worker 3: CHECK 4-7 (check the release notes)
• dependencies: BUILD -> TEST, TEST -> RELEASE, WRITE -> CHECK, CHECK -> RELEASE
• deadline 12 minutes, makespan 12 minutes, ON TIME

There are two parallel chains. BUILD -> TEST proves the build, and WRITE -> CHECK prepares and checks the release notes. Both must finish before RELEASE, the final publish step. The violet wires show those dependencies.

It is one representative use case, not a diagram of every task family. The table above is the full scope.

Use cases

• CI and build scheduling. Assign build, test, documentation and release jobs to runners with time budgets and dependencies, then check the makespan against a deadline.
• Local batch inference. Pack jobs onto workers or GPU slots without exceeding a worker's capacity, and return the assignment.
• Dependency-constrained pipelines. Place jobs on a timeline with before constraints and return worker plus start time for every job.
• Dispatch validation. Check a proposed assignment against worker capacity before it runs, and answer yes/no.
• Urgent triage. When one worker is free and two jobs are waiting, pick the one with the nearer deadline (for example, a live rollback before a background re-index).

Dispatcher returns a plan for the caller to execute; it does not run the jobs or act as a general chat model.

Evaluation

Held-out Hive role suites, greedy decoding:

role	path	accuracy	n	median
dispatcher	HF	0.950	40	1.082 s
dispatcher	GGUF	0.925	40	0.275 s

Reproduce

The held-out task files, deterministic checkers and endpoint evaluator are in the Hive harness. The JSON files under eval/ are the exact evaluation records: greedy decoding, response_format enabled for JSON-shaped tasks, and Q4_K_M for the GGUF row. Weight hashes are in SHA256SUMS.

Training

• Base: LiquidAI/LFM3-1.7B-Base
• Method: TRL SFT + LoRA (r=16, alpha=32, dropout 0.05)
• Data: synthetic, checker-validated Hive role rows; no user data
• Format: LFM3 chat template with enable_thinking=false; assistant-only loss
• Run: 1800 rows, 2.0 epochs, 226 steps
• Peak VRAM: 8.77 GiB on one 20 GB AMD card

Usage

Merged transformers model:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "SkyIsNotGreen/hive-dispatcher-1.7b-v1"
model = AutoModelForCausalLM.from_pretrained(
    repo, subfolder="model", dtype="bfloat16", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained(
    repo, subfolder="model", trust_remote_code=True)

The merged model is the supported checkpoint; this release does not ship a standalone adapter.

GGUF with llama.cpp:

llama-server -m gguf/hive-dispatcher-1.7b-v1-Q4_K_M.gguf --alias dispatcher-1.7b \
  --host 127.0.0.1 --port 8163 -ngl 999 -c 8192 -np 2 \
  --jinja --reasoning off --no-webui

--reasoning off is required for this LFM3 template.

Run with Hive

hive serve can fetch this GGUF and start it as a role tier; see the Hive install guide. The direct llama-server command above is the minimal portable path.

Limitations

Dispatcher-only release of the Operator v2 adapter. It was initialized from Hive Conductor and trained with dispatcher+agency data; the dispatcher gate is the release gate. Agency is served by Hive Tool Manager.

License and attribution

This is a derivative of LiquidAI/LFM3-1.7B-Base, released under the LFM Open License v1.0 (see LICENSE). The base model is by Liquid AI. Commercial use by an entity above USD 10M annual revenue is not licensed under LFM 1.0.

Modified files: merged weights and GGUF conversion. See NOTICE.