LFM2.5-230M Propella

LFM2.5-230M Propella is a compact, high-efficiency language model designed specifically to generate structured text annotations in JSON format. It was distilled from larger annotation models to provide fast, lightweight classification and metadata extraction for English text.

🛠️ Model Details

• Architecture: Small Language Model (SLM) / Distilled Model (~230M parameters)
• Language: English
• Task: Structured JSON text annotation and metadata extraction
• Base Model / License: LFM Open License v1.0 (inheriting from the base model)
• Training Dataset: agentlans/c4-en-propella (annotated using ellamind/propella-1-0.6b)

🚀 Quick Start / Usage

Here is how you can use the model with the Hugging Face transformers library for text annotation:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "agentlans/LFM2.5-230M-Propella"

# Load tokenizer and model with automatic dtype optimization
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    device_map="auto", 
    torch_dtype="auto"
)

# Use the model's chat template for proper instruction formatting
messages = [
    {
        "role": "user", 
        "content": "Annotate as JSON:\n<text>Your text to annotate goes here.</text>"
    }
]

# Apply chat template with generation prompt enabled
inputs = tokenizer.apply_chat_template(
    messages, 
    add_generation_prompt=True,  # Crucial for chat model generation
    return_tensors="pt", 
    return_dict=True
).to(model.device)

# Generate response tokens
outputs = model.generate(
    **inputs, 
    max_new_tokens=512,
    do_sample=True,          # Optional: enables sampling for more creative generation
    temperature=0.3          # Optional: lower temperature keeps structured outputs like JSON more stable
)

# Decode only the newly generated tokens (excluding the input prompt)
input_len = inputs["input_ids"].shape[1]
generated_tokens = outputs[0][input_len:]
print(tokenizer.decode(generated_tokens, skip_special_tokens=True))

📝 Example Annotations

• The input prompt must be formatted as Annotate as JSON:\n<text>___</text>
• The outputs below have been pretty-printed

Source:

Popular-Wrangler1584. (2026). Found this on the ground during a walk last winter. Dated 7/31/2020 [Online forum post]. Reddit. https://www.reddit.com/r/FoundPaper/comments/1w4shm7/found_this_on_the_ground_during_a_walk_last/

Input:

Annotate as JSON:
<text>7/31/2020
It is insane to believe the thoughts in your head.
One of the primary reasons we suffer is because we believe what we think
Why?
Why do we believe the thoughts in our head?
We don't believe the thoughts in others' heads when they speak them to us.
When we read a book - which is nothing but the recording of somebody else's thoughts - we can take them or leave them.
But why is it that we are so prone to grasp at the thoughts that occur within our own mind - to hold onto them and become identified with them?
We don't seem to be able to put them down even when the cause suffering!</text>

Output:

{
  "content_integrity": "complete",
  "content_ratio": "complete_content",
  "content_length": "brief",
  "one_sentence_description": "A philosophical question asks why we hold onto thoughts of others, despite not believing them.",
  "content_type": [
    "qa_structured",
    "opinion_editorial"
  ],
  "business_sector": [
    "general_interest"
  ],
  "technical_content": [
    "non_technical"
  ],
  "information_density": "dense",
  "content_quality": "good",
  "audience_level": "general",
  "commercial_bias": "none",
  "time_sensitivity": "evergreen",
  "content_safety": "safe",
  "educational_value": "basic",
  "reasoning_indicators": "basic_reasoning",
  "pii_presence": "no_pii",
  "regional_relevance": [
    "culturally_neutral"
  ],
  "country_relevance": [
    "none"
  ]
}

Source:

CNN. (2026, September 22). Trump to address United Nations General Assembly. CNN. https://edition.cnn.com/2026/09/22/politics/trump-un-general-assembly-speech

Input:

Annotate as JSON:
<text>The world feels like it’s spinning out of control.

The planet’s two major wars, in Iran and Ukraine, are no longer contained. Experts warn of doomsday scenarios in which AI could wipe out humanity. Soaring energy prices rock industrialized economies blighted by inequality. Floods, fires and blistering heat are turning the climate crisis biblical.

Meanwhile, global institutions that bankrolled relative stability crumble. The United Nations, where world leaders gather this week, looks inept as strongmen spurn its core values: democracy, human rights and international aid.

At any other time over the last 80 years, the US would serve to restore stability. But remarkably, the West’s superpower is now an agent of global volatility.

Step forward, Mr. President.

Donald Trump’s annual address to the UN General Assembly on Tuesday comes with his popularity plunging ahead of critical midterm elections and as he’s stuck in the kind of war he vowed never to fight.

Yet Trump is expected to argue that the US has never been as mighty or esteemed. Officials say he’ll boast of spreading global peace and tackling crises his predecessors ducked.

“Our country is respected again. We’re respected like never before. I go to meetings and everyone stands up. They didn’t stand up for our last president; they didn’t care about him,” Trump said earlier this month.

Much of the world disagrees.</text>

Output:

{
  "content_integrity": "complete",
  "content_ratio": "complete_content",
  "content_length": "brief",
  "one_sentence_description": "Opinion piece arguing that the US presidential election is a political anomaly, contrasting with global stability.",
  "content_type": [
    "opinion_editorial"
  ],
  "business_sector": [
    "government_public"
  ],
  "technical_content": [
    "non_technical"
  ],
  "information_density": "moderate",
  "content_quality": "adequate",
  "audience_level": "general",
  "commercial_bias": "none",
  "time_sensitivity": "time_sensitive",
  "content_safety": "safe",
  "educational_value": "minimal",
  "reasoning_indicators": "basic_reasoning",
  "pii_presence": "no_pii",
  "regional_relevance": [
    "north_american",
    "european",
    "middle_eastern"
  ],
  "country_relevance": [
    "united_states",
    "iran",
    "ukraine"
  ]
}

Source:

Ma, Y., Wang, H.-M., He, J., & Yan, Q. (2026). Two-stage refinement sparse channel estimation for reconfigurable intelligent metasurface antenna (RIMSA) massive MIMO. arXiv. https://doi.org/10.48550/arXiv.2609.22756

Input:

Annotate as JSON:
<text>Two-Stage Refinement Sparse Channel Estimation for Reconfigurable Intelligent Metasurface Antenna (RIMSA) Massive MIMO
Yakun Ma, Hui-Ming Wang, Jiaping He, Qingli Yan

    To meet the increasing demands for high data rates and large capacity, next generation wireless communication systems require transceivers equipped with a large number of antennas. Massive multiple-input multiple-output (MIMO) with metasurface antennas has emerged as a promising solution. In this paper, we investigate the channel estimation problem for the emerging reconfigurable intelligent metasurface antenna (RIMSA) array systems. Specifically, we develop a two-stage refinement (TSR) channel estimation method based on the compressed sensing (CS) principle. In the first stage, we exploit the antenna structure of RIMSA to receive pilots by setting identical phase response vectors across all RIMSAs. In this manner, the coherence of the measurement matrix under the CS framework is reduced and the channel estimation performance is improved. However, this special design introduces channel direction-of-arrival (DoA) estimation ambiguity and yields an ambiguous candidate DoA set. In the second stage, we optimize the phase responses of the metamaterial elements to resolve the ambiguity and accurately estimate the DoAs and channel coefficients. Overall, the estimation accuracy is improved in the first stage at the cost of ambiguity, and this ambiguity is eliminated in the second stage. We illustrate the performance advantages of the TSR method by presenting the numerical results and comparing it with the existing methods.</text>

Output:

{
  "content_integrity": "complete",
  "content_ratio": "complete_content",
  "content_length": "brief",
  "one_sentence_description": "Abstract of a research paper investigating two-stage refinement techniques for channel estimation in reconfigurable intelligent metasurface antenna arrays.",
  "content_type": [
    "analytical"
  ],
  "business_sector": [
    "academic_research",
    "telecommunications",
    "hardware_electronics"
  ],
  "technical_content": [
    "scientific",
    "math_heavy"
  ],
  "information_density": "dense",
  "content_quality": "excellent",
  "audience_level": "expert",
  "commercial_bias": "none",
  "time_sensitivity": "slowly_changing",
  "content_safety": "safe",
  "educational_value": "basic",
  "reasoning_indicators": "analytical",
  "pii_presence": "no_pii",
  "regional_relevance": [
    "culturally_neutral"
  ],
  "country_relevance": [
    "none"
  ]
}

Source:

Norris, M. (2026, March 29). Easy homemade chocolate chip cookies. Melissa K. Norris. https://melissaknorris.com/easy-homemade-chocolate-chip-cookies/

Input:

Annotate as JSON:
<text>Easy Homemade Chocolate Chip Cookies (No Shortening, Whole Ingredients, 5 Dozen)
Melissa Norris
These easy homemade chocolate chip cookies have been in my family for nearly 20 years. They're made with real, whole-food ingredients and without a drop of shortening or hydrogenated oil. Soft and chewy with a slightly crisp edge, they come together in about 20 minutes with pantry staples your great-grandmother would recognize. The recipe makes about 60 cookies, freezes beautifully as baked cookies or raw dough, and works with all-purpose, soft white wheat, or spelt flour with simple adjustments. No artificial vanilla, no Crisco, just a classic cookie worth passing down.
5 from 9 votes
Print Recipe
Pin Recipe
Prep Time 10 minutes mins
Cook Time 8 minutes mins
Resting Time 2 minutes mins
Total Time 20 minutes mins
Course Dessert, Snack
Cuisine American
Servings 60 cookies
Calories 116 kcal
Equipment
    ▢
    Stand Mixer
    ▢
    Cookie Baking Sheet
    ▢
    Cookie Scoop optional
    ▢
    Spatula
    ▢
    Cooling Rack

    ▢ 1.5 cups unsalted butter if using salted, reduce the salt to ¾ teaspoon
    ▢ ¾ cup brown sugar
    ▢ ¾ cup granulated sugar
    ▢ 2 large eggs room temperature
    ▢ 2 teaspoons vanilla extract
    ▢ 3 cups all-purpose flour sifted
    ▢ 1 teaspoon baking soda
    ▢ 1 teaspoon salt
    ▢ 2 cups chocolate chips

Instructions
    Before You Begin: Preheat the oven to 375°F.
    Step 1: In a large bowl, cream together the softened butter, brown sugar, and granulated sugar until light and fluffy, about 3 to 5 minutes. Beat in the eggs and vanilla extract.
    Step 2: Add 2 ½ cups of the sifted flour to the bowl along with the baking soda and salt. Mix just until combined.
    Step 3: If the dough feels too sticky, add the remaining ½ cup of flour. If not, skip that step. The dough should hold its shape but still be soft.
    Step 4: Stir in the chocolate chips until evenly mixed.
    Step 5: Drop rounded tablespoonfuls of dough onto the cookie sheet, spacing them 2 to 3 inches apart.
    Pro Tip: Using a small cookie scoop helps keep them uniform.
    Step 6: Bake for 8 to 10 minutes, until the tops are just barely browned. The cookies will still look soft but will firm up as they cool.
    Step 7: Remove from the oven and let the cookies rest on the baking sheet for 5 minutes before transferring them to a cooling rack.
    Step 8: Enjoy with a glass of milk, or transfer to a freezer-safe container and freeze for up to 2 months. Store leftovers in an airtight container at room temperature for 5-7 days.

Notes
Baking Tips:

    Make sure your butter is softened, not melted. Soft butter gives the best texture with tender centers and crisp edges.
    Use room-temperature eggs so they blend smoothly with the butter and sugar. If your eggs are cold, run them under warm water for a minute.
    Sift the flour to prevent lumps and give the cookies a finer crumb.
    Start with 2 ½ cups of flour, then add up to ½ cup more only if the dough is sticky. This helps you adjust for humidity and altitude.
    Fresh-Milled Flour: Soft white wheat or spelt work best for cookies. For soft white wheat (and any fresh-milled flour other than spelt): reduce flour by 3 tablespoons per cup. For spelt: increase flour by 2 to 4 tablespoons per cup. After mixing, rest the dough 15 minutes before scooping to allow full absorption before assessing consistency and deciding whether to add the remaining half cup of flour.
    Do not overbake. The cookies should look slightly underdone when you remove them from the oven. They will continue to bake on the hot pan as they rest.
    For gifting or storage, cool cookies completely and stack between parchment paper in a tin or airtight container.

Storage and Make-Ahead Ideas

    Store baked cookies in an airtight container at room temperature for up to 5 days.
    Freeze baked cookies in a freezer-safe container for up to 3 months. Rewarm in a 300°F oven for 5 minutes to restore that fresh-baked texture.
    Freeze extra dough by scooping it into balls and freezing on a tray. Once solid, store in a freezer bag. Bake from frozen, adding 1 to 2 minutes to the baking time.

Nutrition
Serving: 1cookieCalories: 116kcalCarbohydrates: 14gProtein: 1gFat: 7gSaturated Fat: 4gPolyunsaturated Fat: 0.2gMonounsaturated Fat: 1gTrans Fat: 0.2gCholesterol: 18mgSodium: 61mgPotassium: 31mgFiber: 0.2gSugar: 9gVitamin A: 151IUCalcium: 10mgIron: 0.3mg</text>

Output:

{
  "content_integrity": "complete",
  "content_ratio": "complete_content",
  "content_length": "brief",
  "one_sentence_description": "Recipe for homemade chocolate chip cookies with tips on baking, storage, and nutrition.",
  "content_type": [
    "procedural",
    "instructional",
    "reference"
  ],
  "business_sector": [
    "food_beverage_hospitality",
    "general_interest"
  ],
  "technical_content": [
    "non_technical"
  ],
  "information_density": "dense",
  "content_quality": "good",
  "audience_level": "general",
  "commercial_bias": "none",
  "time_sensitivity": "evergreen",
  "content_safety": "safe",
  "educational_value": "moderate",
  "reasoning_indicators": "minimal",
  "pii_presence": "no_pii",
  "regional_relevance": [
    "north_american"
  ],
  "country_relevance": [
    "united_states"
  ]
}

Source:

Delhi Metro Rail Corporation. (n.d.). Delhi Metro map. https://delhimetrorail.com/map

Input:

Annotate as JSON:
<text>05:37 AM - 05:56 AM 19 Mins

    Fare*
    ₹ 53
    Stations
    7

* This is a single journey token fare. Use smart card to avail discounts. View More

*Time shown is estimated travel time only (subject to change due to any inevitable reasons). Passengers are advised to keep extra time to travel.

    DWARKA SECTOR - 12
    11 Mins
    Platform No. 2
    Towards DWARKA SECTOR - 21
    Show all 4 stations
    DWARKA SECTOR - 21

Change here Move Towards orange line
5 Mins

    DWARKA SECTOR - 21
    3 Mins
    Platform No. 4
    Towards NEW DELHI (Yellow & Airport Line)
    AIRPORT (T-3)</text>

Output:

{
  "content_integrity": "complete",
  "content_ratio": "complete_content",
  "content_length": "brief",
  "one_sentence_description": "A fare booking page for a travel journey token in the DWARKA SECTOR in India.",
  "content_type": [
    "transactional",
    "structured_data"
  ],
  "business_sector": [
    "travel_aviation",
    "travel_aviation"
  ],
  "technical_content": [
    "non_technical"
  ],
  "information_density": "dense",
  "content_quality": "adequate",
  "audience_level": "general",
  "commercial_bias": "pure_marketing",
  "time_sensitivity": "time_sensitive",
  "content_safety": "safe",
  "educational_value": "none",
  "reasoning_indicators": "none",
  "pii_presence": "no_pii",
  "regional_relevance": [
    "south_asian"
  ],
  "country_relevance": [
    "india"
  ]
}

⚙️ Training Configuration

• Epochs: 1.0 (trained on the full dataset)
• Context Length: 2048 cutoff length
• Optimization: NEFTune (α = 5), neat sequence packing
• LoRA Parameters: Rank 32, α = 64, RSLoRA, dropout 0.8

⚠️ Limitations & Considerations

• JSON Validity: Because small language models can occasionally break formatting syntax, it is recommended to wrap your inference pipeline with a JSON validator/parser (such as json_repair or Pydantic) to handle edge cases.
• Performance Trade-offs: While highly efficient, annotations may not match the absolute performance or nuance of the larger teacher model (ellamind/propella-1-0.6b).
• Language Support: Trained exclusively on English text; performance on other languages is not guaranteed.

📜 Licence

Distributed under the LFM Open License v1.0, matching the terms of the base model.