chore: init

.gitattributes

@@ -33,3 +33,16 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.IQ4_XS.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.IQ2_XXS.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.Q2_K_L.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.Q3_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.Q6_K.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.IQ2_M.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.BF16.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.Q5_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.F16.gguf filter=lfs diff=lfs merge=lfs -text
+gemma-3-1b-it.IQ3_XXS.gguf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,77 @@
 ---
-license: mit
+license: gemma
+base_model: unsloth/gemma-3-1b-it-unsloth-bnb-4bit
+library_name: llama.cpp
+pipeline_tag: text-generation
+tags:
+  - gemma
+  - gemma-3
+  - gguf
+  - text-generation
+  - telegram
+  - trading-signals
+  - structured-output
+  - json
 ---
+
+# Telegram Signal Extract — Gemma 3 1B
+
+A fine-tuned **Gemma 3 1B Instruct** model for extracting a compact trading-signal payload from Telegram channel messages. It is intended to return either a payload such as:
+
+```json
+{"symbol":"ONEUSDT","side":"BUY"}
+```
+
+or the literal JSON value:
+
+```json
+null
+```
+
+The model was trained with supervised examples pairing Telegram messages with a signal payload or `null`. For best observed output, use **temperature 0** and the task system prompt below. Fine-tuning alone does not guarantee valid JSON or correct classifications; validate model output before using it.
+
+## System prompt
+
+```text
+You extract trade signals from Telegram channel messages. For an actionable open/close signal, reply with ONLY a JSON object like {"symbol": "BTCUSDT", "side": "BUY"} where symbol is the uppercase trading pair and side is BUY or SELL. If the message is not an actionable signal, reply with exactly: null
+```
+
+## Files
+
+| File | Quantization / precision | Approx. size |
+|---|---:|---:|
+| `gemma-3-1b-it.Q4_K_M.gguf` | Q4_K_M | 769 MiB |
+| `gemma-3-1b-it.Q5_K_M.gguf` | Q5_K_M | 812 MiB |
+| `gemma-3-1b-it.Q6_K.gguf` | Q6_K | 965 MiB |
+| `gemma-3-1b-it.Q8_0.gguf` | Q8_0 | 1020 MiB |
+| `gemma-3-1b-it.F16.gguf` | F16 | 1.87 GiB |
+| `gemma-3-1b-it.BF16.gguf` | BF16 | 1.87 GiB |
+| Other `IQ*` and `Q2`/`Q3` GGUFs | Smaller quantizations | See file sizes on the Files tab |
+
+The repository is large because it includes **all exported variants**. For most local use, start with `Q4_K_M`; use a higher precision variant if you have enough memory and want to compare quality.
+
+## Task and output
+
+- **Input:** one raw Telegram message.
+- **Positive output:** `{"symbol":"<UPPERCASE_PAIR>","side":"BUY"}` or `{"symbol":"<UPPERCASE_PAIR>","side":"SELL"}`.
+- **Negative output:** `null`.
+- The payload contains `symbol` and `side` only. It does not estimate trade size, profitability, or risk.
+
+## Training
+
+- **Method:** QLoRA (4-bit base model + LoRA adapter, rank 16 / alpha 16 over the attention and MLP projections), later merged and converted to GGUF.
+- **Base checkpoint:** `unsloth/gemma-3-1b-it-unsloth-bnb-4bit`.
+- **Data:** the same Telegram message → payload-or-`null` task as the Llama 3.2 1B and Qwen3 0.6B variants of this model (81,339 training examples, 31,846 held-out).
+- **Steps reached:** 70 optimizer steps — approximately **0.25 of one epoch**, so this is a partially trained model, not a converged one.
+- **Eval loss:** **0.0432** at the final checkpoint, best **0.0368** at ~0.20 epoch, measured on 31,846 held-out examples.
+- **Final training loss:** 0.0276 (last logged step).
+
+**Read the eval loss carefully.** It is teacher-forced token loss, not an extraction-accuracy score: a low value means the model is very good at continuing the expected answer, not that it labels messages correctly. Use task metrics (exact match on `symbol` + `side`, plus precision/recall on signals vs `null`) before trusting it.
+
+## Intended use and limitations
+
+This is an experimental information-extraction model, not a trading system or investment recommendation. It can miss signals, misread tickers/direction, or return invalid output. **Do not automatically place orders from its responses.** Parse the output as JSON, reject anything outside the expected schema, and independently validate symbols and trading actions.
+
+## Base model and license
+
+This model is derived from `unsloth/gemma-3-1b-it-unsloth-bnb-4bit`, a 4-bit quantization of `google/gemma-3-1b-it`. Gemma is **not** Apache-2.0 or MIT: it is governed by the **Gemma Terms of Use** and the **Gemma Prohibited Use Policy**, so this model is marked `gemma`. If you distribute it, pass the Gemma Terms of Use along and ship the notice text file Google requires — *"Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms"*. See the [Gemma 3 1B model page](https://huggingface.co/google/gemma-3-1b-it).

gemma-3-1b-it.BF16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e72468a1c781fd6599b6eefe6a87ea35e47c76b8b8b1e22875d85fe043379ba3
+size 2006573440

gemma-3-1b-it.F16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a40e7246b7d9eb78867653355c451c7bc0be3293a5f8100a1d78ab44562e17a8
+size 2006573792

gemma-3-1b-it.IQ2_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cbf31d7438dfdc5c30b79139bbe61d867c4f8748d8c0684d66192792913b318
+size 669783776

gemma-3-1b-it.IQ2_XXS.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47432737eb53a1314ad2239b075d97ca126f261efa0c5a37008686e4c27422c8
+size 650639840

gemma-3-1b-it.IQ3_XXS.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d7035f2e1f0e6aa3f2236c23c4aa78fcd9692a36a9f7fc3e7a28ea8047e3961
+size 680110304

gemma-3-1b-it.IQ4_XS.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c46846c0d1b1949120eb865fb5d710d33429ca8f125ae330cd3deeaf80f5144e
+size 714435296

gemma-3-1b-it.Q2_K_L.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e9f762e1558e728e847f94bd5e4940f288a801e3b1e6ecf3023da84fea09a0a
+size 689814752

gemma-3-1b-it.Q3_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b68bee51a62c37812be6d3a04f5d01b4355a64e655f02efce2b5e3ecb055051
+size 722416352

gemma-3-1b-it.Q4_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8486cfc5c6847356888f2f74b698b4673497f79eb7f911693ea86c23a23b63c
+size 806058464

gemma-3-1b-it.Q5_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce7082b75528981b7047172890cd1b42c1b13ac78147759cc680e39c47a31306
+size 851345888

gemma-3-1b-it.Q6_K.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfa651f05994fe81a62f6c4b143b7855d7d1770d7e22ac58a223c57cddea4b24
+size 1011738848

gemma-3-1b-it.Q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82a6efb23e25671aa8e8b0712ceda588ad2b163c75bec925b89ec79ba34165bf
+size 1069306592