upload README.md

README.md

@@ -1,18 +1,14 @@
 ---
 language:
 - en
-- zh
 license: apache-2.0
 tags:
 - transformers
 - safetensors
-- llama.cpp
 - gguf
 - peft
 - qlora
 - reasoning
-- math
-- code
 base_model:
 - Nanbeige/Nanbeige4.1-3B
 library_name: transformers
@@ -21,47 +17,27 @@ pipeline_tag: text-generation
 
 # EmberForge-3B-Reasoner
 
-EmberForge-3B-Reasoner is a private finetuned Nanbeige 4.1 3B reasoning model release by `strykes`.
+Private finetuned Nanbeige4.1-3B reasoning release by `strykes`.
 
-## What is included
+## Included Artifacts
 
-This repo intentionally includes multiple artifact types:
-
-- **Merged full model (Safetensors)** at repo root (for Transformers / benchmark pipelines)
-- **LoRA adapter** in `adapter/`
-- **GGUF quants** in `gguf/`:
+- Merged full model (Safetensors) at repo root for HF benchmarking
+- LoRA adapter in `adapter/`
+- GGUF in `gguf/`:
   - `Nanbeige4.1-3B-Q5_K_M.gguf`
   - `Nanbeige4.1-3B-Q4_K_M.gguf`
   - `Nanbeige4.1-3B-f16.gguf`
+- Optional archive in `archives/`
 
-## Training summary
-
-- Base model: `Nanbeige/Nanbeige4.1-3B`
-- Method: QLoRA with Unsloth, merged to full weights
-- Dataset: synthetic reasoning instruction dataset (`3500` samples)
-- Epochs: `2`
-- Effective batch size: `16` (batch 1 x grad acc 16)
-- Max sequence length: `4096`
-- Learning rate: `1e-4` with cosine schedule
-- Final reported training loss: `~1.28`
-
-## Quick usage (Transformers)
-
-```python
-from transformers import AutoTokenizer, AutoModelForCausalLM
-
-model_id = "strykes/emberforge-3b-reasoner"
-
-tok = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
-```
-
-## Quick usage (llama.cpp)
+## Training Snapshot
 
-Use files in `gguf/`, e.g. `Q5_K_M` for stronger quality or `Q4_K_M` for lower RAM.
+- Base: `Nanbeige/Nanbeige4.1-3B`
+- Method: Unsloth QLoRA -> merged weights
+- Data: ~3.5k synthetic reasoning samples
+- Epochs: 2
+- Sequence length: 4096
 
 ## Notes
 
-- This is a finetuned model intended for research/benchmarking.
-- Follow upstream Nanbeige license and applicable usage policies.
-- Outputs can still contain errors; validate for critical tasks.
+- Intended for research and benchmarking.
+- Validate outputs before critical use.