Upload README.md with huggingface_hub

README.md

@@ -1,27 +1,28 @@
 ---
+license: other
+license_name: lfm-1.0
+license_link: https://huggingface.co/LiquidAI/LFM2.5-350M/blob/main/LICENSE
 base_model: LiquidAI/LFM2.5-ColBERT-350M
+tags:
+  - colbert
+  - retrieval
+  - multi-vector
+  - late-interaction
+  - gguf
+  - crispembed
+  - ggml
 language:
   - en
+  - es
   - de
   - fr
-  - es
   - it
   - pt
-  - nl
-  - pl
-  - ru
+  - ar
+  - sv
+  - "no"
   - ja
-  - zh
-license: other
-license_name: lfm1.0
-license_link: https://huggingface.co/LiquidAI/LFM2.5-ColBERT-350M/blob/main/LICENSE
-tags:
-  - gguf
-  - colbert
-  - multi-vector
-  - retrieval
-  - text-embeddings-inference
-  - crispembed
+  - ko
 ---
 
 # LFM2.5-ColBERT-350M — CrispEmbed GGUF
@@ -32,71 +33,62 @@ Multi-vector (ColBERT-style) retrieval: per-token embeddings projected to 128 di
 
 **Format note:** These GGUFs use CrispEmbed's internal tensor naming (`lfm.*` prefix, arch=`lfm2`). They include the `colbert.projection.weight` tensor from the `1_Dense` module. **Not** compatible with llama.cpp.
 
----
-
-## Files
+## Model variants
 
-| File | Size | Compression | ColBERT cos vs fp16 |
-|------|------|-------------|---------------------|
-| `lfm2-colbert-f32.gguf` | 677 MB | 1.0x | **0.999995** |
-| `lfm2-colbert-q8_0.gguf` | 361 MB | 1.9x | **0.998** |
-| `lfm2-colbert-q5_k.gguf` | 258 MB | 2.6x | **0.977** |
-| `lfm2-colbert-q4_k.gguf` | 224 MB | 3.0x | **0.959** |
+| File | Quant | Size | ColBERT cos vs F32 |
+|------|-------|------|--------------------|
+| `lfm2-colbert-f32.gguf` | F32 | 677 MB | 0.999995 |
+| `lfm2-colbert-q8_0.gguf` | Q8_0 | 361 MB | 0.998 |
+| `lfm2-colbert-q5_k.gguf` | Q5_K | 258 MB | 0.977 |
+| `lfm2-colbert-q4_k.gguf` | Q4_K | 224 MB | 0.959 |
 
-**Recommendation:** Use Q8_0 for best accuracy/size trade-off. The LFM2 hybrid conv+attention architecture is sensitive to aggressive quantization (similar to NomicBERT SwiGLU layers). Q5_K/Q4_K produce usable vectors but with noticeable degradation.
+## Architecture
 
-## Model
-
-- **Architecture**: 16-layer hybrid (10 ShortConv + 6 GQA attention), hidden=1024
-- **Output**: Per-token embeddings projected to **128d** via ColBERT projection head
-- **Scoring**: Late interaction (MaxSim) — sum of per-query-token max cosine similarities
+- **Backbone**: LFM2.5-350M bidirectional hybrid (16 layers: 10 ShortConv + 6 GQA attention, 1024-dim hidden, SwiGLU FFN)
+- **ColBERT head**: Linear(1024, 128) + L2 normalize per token
+- **Scoring**: MaxSim — max over doc tokens of cosine similarity per query token, summed
 - **Parameters**: 350M + 128K projection head
-- **Languages**: 11 (en, de, fr, es, it, pt, nl, pl, ru, ja, zh)
+- **Languages**: EN, ES, DE, FR, IT, PT, AR, SV, NO, JA, KO (11 languages)
 - **Task prefixes**: `"query: "` for queries, `"document: "` for passages
 
-## Usage with CrispEmbed
-
-### CLI
+## Usage
 
 ```bash
 # ColBERT multi-vector encode
-./crispembed -m lfm2-colbert-q8_0.gguf --colbert "What is the capital of France?"
+./crispembed -m lfm2-colbert-q8_0.gguf --colbert "query: what is deep learning?"
 
 # JSON output (per-token vectors)
 ./crispembed -m lfm2-colbert-q8_0.gguf --colbert --json "query: machine learning"
-```
 
-### Python
+# Server
+./crispembed-server --embed lfm2-colbert-q8_0.gguf --port 8080
+curl -X POST http://localhost:8080/colbert/score \
+  -d '{"query": "what is deep learning?", "documents": ["Deep learning is a subset of ML", "The weather is nice"]}'
+```
 
 ```python
-import crispembed
-
-model = crispembed.load("lfm2-colbert-q8_0.gguf")
+from crispembed import CrispVit
 
-# Check ColBERT capability
-assert model.has_colbert()
-print(f"ColBERT dim: {model.colbert_dim()}")  # 128
+model = CrispVit("lfm2-colbert-q8_0.gguf")
+assert model.has_colbert
 
-# Multi-vector encode
-q_vecs = model.encode_multivec("query: What is the capital of France?")
-d_vecs = model.encode_multivec("document: Paris is the capital of France.")
+# Encode multi-vector representations
+query_vecs = model.encode_multivec("query: what is deep learning?")   # (n_tokens, 128)
+doc_vecs = model.encode_multivec("document: Deep learning uses neural networks")
 
-# Late interaction (MaxSim) scoring
-import numpy as np
-sim_matrix = q_vecs @ d_vecs.T  # [n_q_tokens, n_d_tokens]
-score = sim_matrix.max(axis=1).sum()
-print(f"MaxSim score: {score:.4f}")
+# MaxSim scoring
+score = model.maxsim(query_vecs, doc_vecs)
+print(f"Score: {score:.4f}")
 ```
 
-### Rust
-
 ```rust
 use crispembed::CrispEmbed;
 
-let model = CrispEmbed::load("lfm2-colbert-q8_0.gguf")?;
+let mut model = CrispEmbed::new("lfm2-colbert-q8_0.gguf", 4)?;
 assert!(model.has_colbert());
-let vecs = model.encode_multivec("query: hello world")?;
-// vecs: Vec<Vec<f32>> — one 128d vector per token
+
+let query = model.encode_multivec("query: what is deep learning?");
+let doc = model.encode_multivec("document: Neural networks learn representations");
 ```
 
 ## Conversion
@@ -120,4 +112,8 @@ python models/convert-lfm2-embed-to-gguf.py \
 
 ## License
 
-[LFM1.0](https://huggingface.co/LiquidAI/LFM2.5-ColBERT-350M/blob/main/LICENSE) — same as the base model.
+[LFM Open License v1.0](https://huggingface.co/LiquidAI/LFM2.5-350M/blob/main/LICENSE) — same as the base model.
+
+## Credits
+
+Original model by [LiquidAI](https://huggingface.co/LiquidAI). GGUF conversion and inference engine by [CrispEmbed](https://github.com/CrispStrobe/CrispEmbed).