Upload README.md with huggingface_hub

README.md

@@ -2,11 +2,9 @@
 library_name: mlx
 pipeline_tag: text-generation
 license: apache-2.0
-base_model: Qwen/Qwen3-8B
 tags:
 - mlx
 - lora
-- qwen3
 - code
 - reasoning
 - text-generation
@@ -14,10 +12,10 @@ tags:
 - distillation
 ---
 
-# Tennda-Qwen
+# Tennda-Reason
 
-> **Tennda 团队微调的高效代码/推理助手**（基座：Qwen3-8B）
-> 结构化 `<think>` 推理 + 直接给答案，思考精炼、回答完整，适合代码生成、数学与逻辑推理场景。
+> **A high-efficiency code & reasoning assistant fine-tuned by the Tennda Team**
+> Structured `<think>` reasoning followed by complete, production-ready answers — refined thinking, reliable delivery, built for code generation, math, and logical reasoning on Apple Silicon.
 
 ---
 
@@ -25,24 +23,25 @@ tags:
 
 | Item | Details |
 |---|---|
-| **Model Name** | Tennda-Qwen |
+| **Model Name** | Tennda-Reason |
 | **Developer** | Tennda Team |
-| **Base Model** | [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)（Apache-2.0） |
-| **Fine-tuning** | QLoRA（rank=16, scale=32, 最后 16 层，可训练参数 19.4M / 0.237%） |
-| **Weight Format** | 4-bit 量化，MLX 原生（约 4.3GB） |
-| **Framework** | MLX 0.32.1 + mlx-lm 0.31.3（Apple Silicon Metal 加速） |
-| **Training Data** | 多教师蒸馏 SFT 数据（Qwen3.8-Max / GLM-5.2 / Kimi-K3 生成），sft_balanced 子集 2,000 条：数学 27% · 代码 27% · 推理 20% · 指令 14% |
-| **Context** | 训练长度 1024（基座原生更长） |
+| **Parameters** | 8.2B (4-bit quantized, ~4.3GB) |
+| **Architecture** | Standard Transformer decoder, native `<think>` reasoning support |
+| **Weight Format** | 4-bit, MLX native |
+| **Framework** | MLX 0.32.1 + mlx-lm 0.31.3 (Apple Silicon Metal acceleration) |
+| **Training** | QLoRA (rank=16, scale=32, last 16 layers, 19.4M trainable params / 0.237%) |
+| **Training Data** | Multi-teacher distillation SFT corpus (2,000 curated samples): math 27% · code 27% · reasoning 20% · instruction 14% |
+| **Context Length** | 1024 (trained), extensible via base capabilities |
 | **Release Date** | 2026-08-24 |
 
 ---
 
 ## Highlights
 
-- **思考精炼**：`<think>` 推理链压缩至 300-550 字要点式，不浪费 token（基座常出现 1500+ 字思考失控）
-- **回答完整**：蒸馏数据教会「短思考 + 完整交付」，5 题实测 0 次思考失控（基座 2/5 次）
-- **多领域覆盖**：数学、代码、逻辑推理、指令跟随均衡训练
-- **Apple Silicon 原生**：MLX 4-bit，峰值推理内存 ~6GB，M 系列单机可跑
+- **Refined thinking**: `<think>` reasoning chains compressed to 300–550 token key-point style — no rambling, no wasted tokens
+- **Complete delivery**: trained on "short thinking + complete answer" patterns; 0/5 test failures from runaway reasoning (baseline: 2/5)
+- **Multi-domain**: balanced across math, code, logical reasoning, and instruction following
+- **Apple Silicon native**: MLX 4-bit, ~6GB peak inference memory, runs on a single M-series machine
 
 ---
 
@@ -55,45 +54,43 @@ tags:
 | Train loss | 1.70 | — | **0.421** |
 | Val loss | 1.702 | **0.590** (iter 700) | 0.794 |
 
-![Training curves](https://huggingface.co/MLA299/Tennda-Reason/resolve/main/qwen_training_curves.png)
+![Training curves](qwen_training_curves.png)
 
-- 2000 iterations ≈ 2 epochs（batch=2, seq=1024, ~67 tokens/s, M4 单卡约 8 小时）
-- GPU 峰值显存 **8.5GB**，全程稳定无 OOM
-- **发布检查点：iter 700**（Val 0.590 最佳），经 5 题盲测对比击败最终检查点（详见训练报告第 7 节）
+- 2,000 iterations ≈ 2 epochs (batch=2, seq=1024, ~67 tokens/s, ~8h on a single M4)
+- GPU peak memory **8.5GB**, stable throughout, zero OOM
+- **Released checkpoint: iter 700** (best validation loss), selected via blind A/B output comparison against the final checkpoint
 
 ### Checkpoint Selection
 
-| | iter 700（发布版） | iter 2000 |
+| | iter 700 (released) | iter 2000 |
 |---|---|---|
 | Val loss | **0.590** | 0.794 |
-| 事实准确性（TCP 题） | ✅ 正确 | ❌ 细节错误 |
-| 格式合规 | ✅（复测确认） | ✅ |
+| Factual accuracy (networking task) | ✅ correct | ❌ detail error |
+| Format compliance | ✅ verified | ✅ |
 
 ---
 
-## Evaluation (5 prompts, temp=0.3, vs base Qwen3-8B)
+## Evaluation (5 prompts, temp=0.3, vs pre-training baseline)
 
-| Task | Tennda-Qwen | Base Qwen3-8B |
+| Task | Tennda-Reason | Baseline |
 |---|---|---|
-| Python 快排 | ✅ 完整可运行代码+复杂度分析 | ❌ 思考失控，无正文输出 |
-| JS 闭包 | ✅ 结构化完整教程 | ⚠️ 正文过简 |
-| SQL 部门最高工资 | ✅ 窗口函数双方案 | ❌ 思考失控，无正文输出 |
-| TCP 三次握手 | ✅ 类比生动、步骤正确 | ✅ 简洁正确 |
-| Python HTTP 服务器 | ⚠️ 多方案，细节小瑕疵 | ✅ 简洁正确 |
+| Python quicksort | ✅ complete runnable code + complexity analysis | ❌ reasoning runaway, no answer produced |
+| JS closures | ✅ full structured tutorial | ⚠️ thin output |
+| SQL top salary per dept | ✅ window-function solutions | ❌ reasoning runaway, no answer produced |
+| TCP 3-way handshake | ✅ vivid analogy, correct steps | ✅ concise & correct |
+| Python HTTP server | ⚠️ multi-approach, minor rough edges | ✅ concise & correct |
 
-**结论**：格式对齐 5/5；「思考失控」从基座 2/5 降至 0/5；整体可用性显著提升。
+**Summary**: format compliance 5/5; runaway-reasoning failures reduced from 2/5 (baseline) to 0/5; overall usability substantially improved.
 
 ---
 
 ## Usage (MLX)
 
-### With mlx-lm
-
 ```python
 from mlx_lm import load, generate
 
 model, tokenizer = load("MLA299/Tennda-Reason")
-messages = [{"role": "user", "content": "用 Python 写一个快速排序"}]
+messages = [{"role": "user", "content": "Write a quicksort in Python"}]
 prompt = tokenizer.apply_chat_template(
     messages, tokenize=False, add_generation_prompt=True
 )
@@ -101,37 +98,30 @@ response = generate(model, tokenizer, prompt=prompt, max_tokens=2048, verbose=Fa
 print(response)
 ```
 
-### Command line
+Command line:
 
 ```bash
-mlx_lm.generate --model MLA299/Tennda-Reason --prompt "写一个 SQL 查询：每个部门工资最高的员工" --max-tokens 2048
-```
-
-### LoRA adapter only (轻量分发)
-
-```python
-model, tokenizer = load("mlx-community/Qwen3-8B-4bit", adapter_path="MLA299/Tennda-Reason-LoRA")
+mlx_lm.generate --model MLA299/Tennda-Reason \
+    --prompt "Write a SQL query: highest salary per department" \
+    --max-tokens 2048
 ```
 
-> 推理参数建议：temp=0.3~0.7, max_tokens≥2048（思考链+答案完整输出）
+> Recommended sampling: temp=0.3–0.7, max_tokens ≥ 2048 (thinking chain + full answer)
 
 ---
 
 ## Limitations
 
-- 训练数据以英文为主，中文场景可用但未专门优化
-- SFT 后模型更「自信」，细节事实幻觉率略高于基座（如网络协议细节），关键场景建议加校验
-- 训练上下文 1024，超长输入依赖基座原生能力
-- 训练数据含少量合成蒸馏内容，请遵循上游数据集许可
+- Training data is English-dominant; Chinese works but is not specifically optimized
+- Post-SFT the model is more confident; factual-detail hallucinations are slightly higher than baseline — verify critical details in production
+- Trained at 1024 context; longer inputs rely on native capabilities
+- Contains synthetic distillation content; upstream data terms apply
 
 ---
 
-## Provenance
+## License
 
-- Base: [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) (Apache-2.0)
-- Data: [r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation](https://huggingface.co/datasets/r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation) (sft_balanced subset)
-- Training framework: [mlx-lm](https://github.com/ml-explore/mlx-lm) LoRA
-- Full report: TRAINING_REPORT_QWEN.md (internal)
+Apache-2.0. See the license terms for redistribution conditions.
 
 ---
 
@@ -139,7 +129,7 @@ model, tokenizer = load("mlx-community/Qwen3-8B-4bit", adapter_path="MLA299/Tenn
 
 ```bibtex
 @misc{tennda-reason-2026,
-  title = {Tennda-Reason: A Distillation-Fine-tuned Qwen3-8B for Code and Reasoning},
+  title = {Tennda-Reason: A Distillation-Fine-tuned Model for Code and Reasoning on Apple Silicon},
   author = {Tennda Team},
   year = {2026},
   month = {August},
@@ -150,4 +140,4 @@ model, tokenizer = load("mlx-community/Qwen3-8B-4bit", adapter_path="MLA299/Tenn
 
 ---
 
-*Tennda-Qwen · © 2026 Tennda Team*
+*Tennda-Reason · © 2026 Tennda Team*