market-brief-mixtral-lora-v3

Market Brief Adapter (V3) — Mixtral-8x7B LoRA

AutoScientist Challenge 2026 · Market Analysis & News
Ship version: V3 (richer TrueNorth-enriched training path)

Grounded market briefs on multi-source market context.
Headline / article / equity context → structured analyst note (thesis, direction, catalysts, risks, confidence)—numbers only when present in the source text.

---

Why V3 (ship story)

Experiment trail (honest)

We ship V3 for the Market Analysis & News narrative: multi-source, live-enriched training data with a capable open MoE base. V1 remains a valid alternate checkpoint if you prefer max platform WR alone; V3 is the fuller data story.

---

Intended use

---

How to use

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "mistralai/Mixtral-8x7B-Instruct-v0.1"
# After upload: "YOUR_USER/market-brief-mixtral-lora-v3"
adapter_id = "YOUR_USER/market-brief-mixtral-lora-v3"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id, torch_dtype=torch.float16, device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)

instruction = """You are a market analysis assistant. Read the article and produce a structured equity/news brief. Only use facts present in the article. Do not invent prices, dates, or catalysts.

Output exactly these sections:
## Thesis
## Direction
## Horizon
## Catalysts
## Risks
## Key numbers (from text only)
## Confidence
## What would change the view
## Disclaimer"""

user = f"""{instruction}

HEADLINE: Fed holds rates steady; signals data-dependent path
TICKER: SPX
AS_OF: 2026-07-01

ARTICLE:
The Federal Reserve kept its policy rate unchanged...
"""

messages = [{"role": "user", "content": user}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=800, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Local path:

model = PeftModel.from_pretrained(base, "release/model/adapter")

---

Training data (V3)

Dataset hub (companion release): pilot, synthetic seed, TrueNorth-enriched, merged seed, adapted exports—each with README. See release/datasets/.

---

Evaluation note

0.7806 win rate = Adaption AutoScientist preference (adapted vs base).
It is not trading P&L, FinQA exact-match, or filing-classification F1. Reported the same way as other challenge entries, with that caveat.

---

Output schema

## Thesis
## Direction
## Horizon
## Catalysts
## Risks
## Key numbers (from text only)
## Confidence
## What would change the view
## Disclaimer

---

Limitations

• Can omit or over-weight details; verify against source text.
• TrueNorth-derived context can be noisy or stale; not a substitute for primary filings.
• English-first; educational/synthetic scenarios remain part of the mix.
• Preference WR ≠ calibrated accuracy.
• Needs Mixtral base weights + enough GPU to load MoE + adapter.

---

Bias and risks

News and social digests carry source bias. Model may sound authoritative while wrong. Human review required. Not investment advice.

---

Files

---

Citation

Market Brief Adapter V3 (Mixtral-8x7B LoRA).
Adaption AutoScientist Challenge 2026 — Market Analysis & News.
TrueNorth-enriched seed + Adaptive Data + AutoScientist.
Win rate 0.7806 vs base Mixtral (platform preference).

• Challenge: https://adaptionlabs.ai/blog/autoscientist-challenge
• TrueNorth (data tools): https://docs.truenorth.xyz/

---

Disclaimer

Not investment advice. Research / educational use only. No warranty.