Dharma-AI2 min read

Dharma’s 2B Field Extractor scores 0.834 on its field-extraction bench — ahead of Opus 5 at 0.816

Important
Lab claim

How to read this claim

Start here · how to check ranks, prices, and unknowns Model evaluation

Measurement — Dharma’s 2B Field Extractor scores 0.834 field-level F1 on its 546-document bench, against Opus 5 at 0.816. That figure is Dharma’s.
Bill — Cost vs APIs: ~6×–147× lower (Dharma, L40S @$2.24/hr)
Door — Full public release terms and Hub license name are not restated as a single license string in the extractable post body.

On November 4, 2026, Dharma-AI published Dharma Field Extractor, a 2-billion-parameter model that returns typed JSON fields from document images. On Dharma’s 546-document / 1,638-case bench, it scores 0.834 — above Claude Opus 5 at 0.816 and GPT-5.6 Luna at 0.741 on the same suite. Dharma puts self-hosted L40S cost about 6× to 147× below the frontier API runs it measured.

Same three questions

At a glance

Figures from Dharma-AI’s November 4, 2026 Field Extractor post.
Date	November 4, 2026
Parameters	2B (LoRA on DharmaOCR-LITE v1.1)
Field Extractor F1	0.834 (Dharma)
Claude Opus 5	0.816 (Dharma’s run)
GPT-5.6 Luna	0.741 (Dharma’s run)
Documents / cases	546 docs · 1,638 cases
Training examples	43,384
Cost vs APIs	~6×–147× lower (Dharma, L40S @$2.24/hr)

Quick checks

What is Dharma Field Extractor’s score?

Dharma-AI reports 0.834 field-level F1 on its 546-document / 1,638-case extraction bench (November 4, 2026). Claude Opus 5 scored 0.816 and GPT-5.6 Luna 0.741 on Dharma’s runs of the same suite.

Does a 2B model beat Opus on every document task?

No. This lead is on Dharma’s field-extraction bench (typed JSON fields from document images), not a general chat or Index composite. Ask who measured the fields and which request mode (open vs schema-directed).

Where should you start when reading a specialist-vs-frontier claim?

On Minerka’s field card at https://minerka.com/learn/: ask for the measurement, the bill, and what you still cannot call. For Field Extractor, keep Dharma’s self-host L40S cost column separate from frontier API list rates.

Field card — measurement, bill, door

In today’s briefing

All of today

What changed

Dharma shipped a 2B field-extraction checkpoint with a published F1 table that places it above the frontier API configurations it ran on the same document set.

What it introduces

Dharma Field Extractor — 2B parameters, open and schema-directed JSON field extraction from document images.

Who benefits Anyone wiring document intake into a database, comparing OCR-plus-LLM pipelines to a single extraction checkpoint, and anyone who would otherwise call Opus or GPT for every page.

Try this

When a slide says a 2B model “beats Opus on documents,” ask whether that is Dharma’s field-extraction F1 on the 546-document set, which request mode, and whether cost is self-host GPU time or API list. Keep SROIE/FUNSD anecdotes separate from the aggregate 0.834.

What we do not know

Full public release terms and Hub license name are not restated as a single license string in the extractable post body.
Which four frontier configurations besides Opus 5 and GPT-5.6 Luna/Sol sit in the middle of the 1.8–9.3 point gap are summarized rather than fully tabulated in the extractable text.
Self-host $/page depends on batching and latency targets; the $2.24/hour L40S figure is Dharma’s reference, not a vendor quote for every deployment.

Evidence

Primary Field Extractor post. Parameter count, F1 scores, document counts, and cost multiples are from Dharma-AI’s November 4, 2026 Hugging Face blog. Frontier comparisons are Dharma’s runs on Dharma’s field-extraction benchmark.

Independent Scores and cost multiples are Dharma’s. Third parties can try the demo linked from the post; independent reruns of the full 1,638-case suite are not established here.

Before the Dharma-AI link: the three checks (measurement, bill, door). Start here →

Sources

Field Extractor post Primary · November 4, 2026
Field Extractor post
Dharma-AI

Still unclear? · Start here · ranks, prices, unknown Full public release terms and Hub license name are not restated as a single license string in the extractable post body.

The record

What happened

Dharma-AI published Dharma Field Extractor on November 4, 2026 (Hugging Face blog). The checkpoint is a LoRA on DharmaOCR-LITE v1.1 (2 billion parameters) trained on 43,384 examples covering open extraction and schema-directed JSON. The model takes a document image plus either no schema or an exact typed schema and returns structured JSON, with null for missing fields. Dharma’s benchmark covers 546 documents and three request modes (1,638 cases). Scoring is field-level F1 (path-matched, including table position) multiplied by Levenshtein similarity on matched values and by the share of valid JSON outputs. Dharma reports 0.834 for Field Extractor, 0.816 for Claude Opus 5 (closest), and 0.741 for GPT-5.6 Luna (widest gap of 9.3 points among the four frontier configurations tested). Frontier costs used measured tokens at published list rates with reasoning set to low; Dharma’s figure is GPU time on one L40S at $2.24/hour, which Dharma says is about 6× to 147× cheaper than those API runs depending on the model. SROIE and FUNSD examples in the post were held out from training.

Why it matters

Teams still quote a frontier chat model for high-volume invoice and ID extraction. This post puts a named small specialist above Dharma’s own frontier reruns on that job, with a self-host cost column — the habit is to name the extraction bench and who paid for the page, not to treat a general Index score as the extraction answer.

Training mix includes English multi-domain web documents, Brazilian Portuguese internal documents, and a smaller multilingual share.

Open mode invents no schema; schema mode returns exact keys and nulls for phantom fields rather than inventing values, on Dharma’s account.

Dharma’s earlier OCR paper (DharmaOCR-LITE) made a similar specialization claim on Brazilian Portuguese OCR — treat that as a separate measurement from this field-extraction suite.