SkyIsNotGreen/hive-reader-1.7b-v1

Hive Reader 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 mechanics + RAG specialist for the Hive local control plane: read a document or passage, extract the structured fields, or answer from the passage.

This specialist covers the mechanics, rag seat(s) for the Hive local control plane, built from LiquidAI/LFM3-1.7B-Base. It is not a general chat model.

Use cases

• Invoice and receipt extraction. Pull the payee, date, line items, discount, tax and final total from a short invoice or receipt, and return the requested JSON fields.
• Contract date resolution. Resolve the effective delivery or renewal date after an amendment, rather than returning the original date.
• Order and notice fields. Turn a short order, email or business notice into the structured fields a local workflow needs before routing it.
• Passage-grounded answers. Answer a question from a supplied passage, using only the values in that passage; useful for retrieval checks and document QA.
• Batch document cleanup. Process repetitive invoices, orders and notices locally on a small card when the fields are known and the source text is short.

Reader is document-bound: supply the passage or text and ask for the fields or answer. It is not a general chat model.

Evaluation

Held-out Hive role suites, greedy decoding:

role	path	accuracy	n	median
mechanics	HF	1.000	40	0.193 s
mechanics	GGUF	0.975	40	0.080 s
rag	HF	1.000	40	0.061 s
rag	GGUF	1.000	40	0.060 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: 2104 rows, 3.0 epochs, 396 steps
• Peak VRAM: 7.21 GiB on one 20 GB AMD card

Usage

Merged transformers model:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "SkyIsNotGreen/hive-reader-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)

PEFT adapter:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained(
    "LiquidAI/LFM3-1.7B-Base", dtype="bfloat16", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "SkyIsNotGreen/hive-reader-1.7b-v1", subfolder="adapter")

GGUF with llama.cpp:

llama-server -m gguf/hive-reader-1.7b-v1-Q4_K_M.gguf --alias reader-1.7b \
  --host 127.0.0.1 --port 8159 -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

Mechanics closes invoices, dates, and totals; RAG answers from a supplied passage. Both are document-bound. The model is not a general chat model and should be guarded by its checker in production.

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: the LoRA adapter, merged weights, and GGUF conversion. See NOTICE.