everyone is sleeping on this new JSON extraction model!
- open-weights
- matches Gemini Flash 3.5
- while being 3x faster
pulling structured data out of documents usually takes a chain of OCR, an LLM, and extra code to repair broken JSON.
lift, from @datalabto, does all of you give lift a PDF or image along with a JSON schema, and get back a JSON object that matches that schema.
the model reads every page of a document in one pass, so values that span multiple pages still come out right. any document you can describe in a schema works, from invoices to research papers.
here's what you get with lift:
- 90.2% field accuracy on 11,000 fields
- 9B model, nearly matches Gemini Flash 3.5 at 3x the speed
- built not to hallucinate; returns null if missing
- self-hostable with vLLM or Hugging Face
GitHub repo: github.com/datalab-to/lift
(don't forget to star 🌟)
- open-weights
- matches Gemini Flash 3.5
- while being 3x faster
pulling structured data out of documents usually takes a chain of OCR, an LLM, and extra code to repair broken JSON.
lift, from @datalabto, does all of you give lift a PDF or image along with a JSON schema, and get back a JSON object that matches that schema.
the model reads every page of a document in one pass, so values that span multiple pages still come out right. any document you can describe in a schema works, from invoices to research papers.
here's what you get with lift:
- 90.2% field accuracy on 11,000 fields
- 9B model, nearly matches Gemini Flash 3.5 at 3x the speed
- built not to hallucinate; returns null if missing
- self-hostable with vLLM or Hugging Face
GitHub repo: github.com/datalab-to/lift
(don't forget to star 🌟)
Oct 3, 2026 · 12:30 PM UTC
47
120
973
61,053
@akshay_pachaar @datalabto Single-step PDF to schema-validated JSON is huge
0
0
0
139
@akshay_pachaar @datalabto looks abandoned? 4m in AI time is eternity!
0
0
0
625
@akshay_pachaar @datalabto how does it handle deeply nested structures? most extraction models start hallucinating keys past 4-5 levels.
1
0
2
296
@akshay_pachaar @datalabto Cutting out the repair pipeline is a massive win. Honestly, anything that saves me from wrestling with broken JSON chains is a total game changer.
0
0
0
41
@akshay_pachaar @datalabto the ocr then llm then repair chain is where most extraction bugs come from, one model doing all of it would cut a lot
0
0
2
415
@akshay_pachaar @datalabto How does it handle a table that spans two pages of a scanned PDF?
1
0
1
175
@akshay_pachaar @datalabto Leases are the first thing i thought of here, the term up front... the renewal clause pages later. An OCR chain hands those back as separate pages and you match them up yourself afterwards. But one pass over the whole thing and nobody has to do that
0
0
1
183
@akshay_pachaar @datalabto ヒートマップで見る視線誘導も重要ですが、
実は「認知負荷」が売れるデザインのカギ。
脳は情報を絞るほど理解が深まるので、
LPは料理の盛り付けみたいに、
余計な盛り付けを省いて味の本質を際立たせる感じなんです。
実は「認知負荷」が売れるデザインのカギ。
脳は情報を絞るほど理解が深まるので、
LPは料理の盛り付けみたいに、
余計な盛り付けを省いて味の本質を際立たせる感じなんです。
0
0
0
4
@akshay_pachaar @datalabtos://t.co/RDkhvWsuWh, thentityenricher.aie challenges mentioned:
- Uses the PDF input capability of your chosen model directly.
- JSON is validated both by the LLM—by supplying the output schema through the structured-output channel—and by the application.
- Missing information: The schema is split so each LLM call focuses on its own subset of data, and JSON validation allows retries when required information is missing.
- Avoiding hallucinations: Extraction can run across multiple models, with a judge model to arbitrate disagreements.
- Fast extraction: The application lets you run benchmarks on your own use case to choose the models offering the best balance of quality, cost, and speed.
- Uses the PDF input capability of your chosen model directly.
- JSON is validated both by the LLM—by supplying the output schema through the structured-output channel—and by the application.
- Missing information: The schema is split so each LLM call focuses on its own subset of data, and JSON validation allows retries when required information is missing.
- Avoiding hallucinations: Extraction can run across multiple models, with a judge model to arbitrate disagreements.
- Fast extraction: The application lets you run benchmarks on your own use case to choose the models offering the best balance of quality, cost, and speed.
0
0
0
74
@akshay_pachaar @datalabto Nice for docs you already have. For pages still on the open web, we split the job:
Search → shortlist → Extract to main-content markdown → *then* structured parse. Schema extractors hate nav/cookie chrome more than missing OCR.
agentsearchhq.com/extract?utm_so…
Search → shortlist → Extract to main-content markdown → *then* structured parse. Schema extractors hate nav/cookie chrome more than missing OCR.
agentsearchhq.com/extract?utm_so…
0
0
0
17
@akshay_pachaar @datalabto OCR + LLM + JSON repair chain is exactly the pipeline I keep rebuilding. What's the eval here, actual table/field extraction accuracy? 'matches Gemini Flash' on what set, with what schema complexity?
0
0
0
51
@akshay_pachaar @datalabto 一步式抽取最怕错误不可见。结构化数据得能回指到原文 span，不然错一个字段整批脏，下游比手工还贵。lift 的输出带不带 source grounding？不带的话上生产还得自己套一层校验。
0
0
0
201
@akshay_pachaar @datalabto I wonder what are domains where less than 100% accuracy at reading data is ok 🤔
0
0
0
8
@akshay_pachaar @datalabto open weights matching flash-tier on extraction is the sleeper category — documents, websites, it's all just JSON underneath
0
0
0
38
Bring your train dreams to life with exciting 3D tracks, thrilling stunts, and endless creative adventures! Spark imagination and problem-solving skills through fun, hands-on STEM play.
Shop now: kiddibo.com/3d-train
Shop now: kiddibo.com/3d-train
0
19
275
235,158
@akshay_pachaar @datalabto Matching a closed model on a benchmark is not matching it on your data: Clef-flash went 64% vs 78% on 262 of our real decisions. Yet one signal told us exactly which of its answers to keep. Tested lift on your own docs yet? x.com/stas_sorokin_/…
Cloudflare's Clef-flash on 262 of our real production decisions: 64% right. The API we run today: 78%.
Cloudflare open-sourced Clef on 1 Oct as a drop-in for the Jev API. So we took the labeled holdout cases from 17 decision systems we run every day and gave it the same word:
→ Clef-flash 9B, 4-bit, on a 16 GB Mac mini: 168/262
→ Jev on the same cases: 204/262
→ won 1 system, tied 2, lost 14
→ biggest gaps: "is this lead ready to buy" 14/20 vs 20/20, "will this client reply" 7/20 vs 12/20
The twist: its confidence is honest.
→ at 90%+ sure it got 32 of 35 right
→ sure ones on the Mac mini, the rest on Jev: 203/262, one short of Jev alone, and 13% of calls never leave the building
One caveat: those instructions were tuned on Jev for weeks. Clef-flash read them cold.
Would you route your classifier calls to open weights, or only the sure ones?
Repost this now, because "drop-in compatible" will be read as "same answers".
Follow for numbers from our own runs, not the launch post.
0
0
0
32
@akshay_pachaar @datalabto 3x faster is compelling, but I’d want exact-match extraction rates split by document type before calling it a clean win.
0
0
0
18
@akshay_pachaar @datalabto what languages does it support?
0
0
0
58
@akshay_pachaar @datalabto Returning null for missing fields is a useful detail. I’d be curious to see examples with messy tables and values that carry over between pages.
0
0
0
221
@akshay_pachaar @datalabto The JSON repair step is where these setups usually break. If lift removes that whole chain, I'll wire it into a webhook and check the failure rate on malformed payloads — that's the only number that matters for a pipeline.
0
0
0
124
Upgrade vehicle interiors with heavy-duty tactical seat protection built for tough daily drives!
Get it now🎖️theclassicgear.com/seat
Get it now🎖️theclassicgear.com/seat
40
39
460
76,439
@akshay_pachaar @datalabto The speedup is compelling. I’d still separate schema-valid JSON from usable extraction: can the eval score source span, field confidence, and behavior on missing or ambiguous fields? Repairing malformed JSON is the easy failure mode.
1
0
0
235
@akshay_pachaar @datalabto How does it handle a table that runs across two pages?
0
0
0
245
@akshay_pachaar @datalabto 90.2% across 11,000 fields. how many invoices came back with every field right?
1
0
0
425
@akshay_pachaar @datalabto null instead of hallucinating missing fields is the whole sell. repair loops suck
1
0
0
274
@akshay_pachaar @datalabto This single-step solution is really neat, the speed and accuracy are impressive, I'll test it later.
0
0
0
176