GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
flamiinngo
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
flamiinngo
Model Tree
mistralai/Mixtral-8x7B-Instruct-v0.1
Input Modalities
Output Modalities
Supported Functionality
Head-to-head win rate against the base model:
| Evaluation | Base | Adapted |
|---|---|---|
| Training distribution | 28 | 72 |
| Market-Analysis category (all tasks) | 40 | 60 |
Both numbers are wins in a paired comparison, not accuracy percentages. The two figures differ because the category evaluation spans a wider range of tasks than the training data covers — the model is strongest on the task types it was trained on and holds a smaller edge across the category as a whole.
Worth stating plainly: an earlier checkpoint trained on shorter, more templated answers scored 64–36 on the category evaluation while scoring only 45–55 on its own training distribution. Concise answers generalised better; richer answers won more decisively in-distribution. This release is the latter.
Extract the adapter — the archive unpacks flat, so give it its own directory:
bash
mkdir -p market-adaptertar --zstd -xf adaption_mixtral_8x7b_instruc_financial_news_classifie_45d894da.tgz -C market-adapter
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelBASE = "mistralai/Mixtral-8x7B-Instruct-v0.1"ADAPTER = "./market-adapter"tokenizer = AutoTokenizer.from_pretrained(ADAPTER)model = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, ADAPTER)model.eval()messages = [{"role": "user","content": "What area of the market does this headline concern, and why does ""it matter? Citi cuts Microsoft price target, citing foreign ""exchange headwinds",}]inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)out = model.generate(inputs, max_new_tokens=200, do_sample=False)print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Hardware: Mixtral-8x7B is a 46.7B-parameter mixture-of-experts model — roughly 94 GB in bf16, or about 25 GB with 4-bit quantisation. The adapter itself is 54 MB.
Task types it was trained on: topic classification, sentiment assessment, market implication, subject identification, who-it-matters-to, and earnings-call analysis.
mistralai/Mixtral-8x7B-Instruct-v0.1| Parameter | Value |
|---|---|
Rank (r) | 16 |
lora_alpha | 32 |
lora_dropout | 0.0 |
| Target modules | q_proj, k_proj, v_proj, o_proj |
| Peak learning rate | 1e-4 |
| Epoch | Step | Eval loss |
|---|---|---|
| 0.64 | 53 | 1.0593 |
| 1.23 | 102 | 1.0188 |
| 1.82 | 151 | 1.0000 |
| 2.41 | 200 | 0.9922 |
| 3.00 | 249 | 0.9912 |
Training loss fell from 1.414 to 0.876. Evaluation loss was still declining at the end of epoch 3, though the last epoch gained only 0.001 — close to converged, with no sign of overfitting.
flamiinngo/market-news-qa — 10,011 market-analysis Q&A pairs across five task types, median answer 16 words, 85% of answers distinct. Derived from zeroshot/twitter-financial-news-topic (MIT), zeroshot/twitter-financial-news-sentiment (MIT), and lamini/earnings-calls-qa (CC-BY-4.0).
Also on Kaggle: dataset · adapted training data
The adapter is released under Apache 2.0, matching the base model's licence. Note that the training data is CC-BY-4.0 and requires attribution — see the dataset card.
| Schedule | cosine, warmup 0.1 |
| Epochs | 3 (249 steps) |